From 0f8fa8394be658279a455d268194cd2d5a3ee700 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Thu, 28 May 2026 22:03:13 -0700 Subject: [PATCH 01/29] some fixes --- .../datalake/workspace_manager.py | 11 +- src/app/dfSlice.tsx | 28 ++ src/views/DataFormulator.tsx | 34 ++- src/views/DataLoadingChat.tsx | 263 +++++++++--------- src/views/DataThread.tsx | 16 +- src/views/UnifiedDataUploadDialog.tsx | 140 ++++------ src/views/dataLoadingSuggestions.ts | 82 ++++-- tests/backend/data/test_workspace_manager.py | 64 ++++- 8 files changed, 369 insertions(+), 269 deletions(-) diff --git a/py-src/data_formulator/datalake/workspace_manager.py b/py-src/data_formulator/datalake/workspace_manager.py index 679452ca..23e37176 100644 --- a/py-src/data_formulator/datalake/workspace_manager.py +++ b/py-src/data_formulator/datalake/workspace_manager.py @@ -169,6 +169,10 @@ def list_workspaces(self) -> list[dict]: workspace. If a workspace directory lacks this file (legacy), it is auto-repaired via :meth:`_ensure_meta`. + Every workspace directory is listed, including empty + "Untitled Session" entries from data-loading chats. Users + manage (rename/delete) these themselves via the sidebar. + Returns list of {"id": str, "display_name": str, "updated_at": str}. """ workspaces = [] @@ -184,13 +188,16 @@ def list_workspaces(self) -> list[dict]: except Exception: continue + tc = meta.get("tableCount") + cc = meta.get("chartCount") + workspaces.append({ "id": child.name, "display_name": meta.get("displayName", child.name), "created_at": meta.get("createdAt") or meta.get("updatedAt"), "updated_at": meta.get("updatedAt"), - "table_count": meta.get("tableCount"), - "chart_count": meta.get("chartCount"), + "table_count": tc, + "chart_count": cc, }) workspaces.sort(key=lambda w: w.get("updated_at") or "", reverse=True) diff --git a/src/app/dfSlice.tsx b/src/app/dfSlice.tsx index a3fe5add..89b075ab 100644 --- a/src/app/dfSlice.tsx +++ b/src/app/dfSlice.tsx @@ -210,6 +210,17 @@ export interface DataFormulatorState { * Transient — not persisted. */ dataLoadingChatResetCounter: number; + /** + * Pending submission queued for the data-loading chat. Set by any + * surface that wants to hand a prompt off to the chat (the menu + * agent input box, suggestion auto-run, external dialog callers). + * `DataLoadingChat` consumes it on render: it clears the slot and + * sends the carried payload as a fresh user message. Using a single + * redux slot (instead of props + a reset counter) eliminates the + * cross-tick race where the parent's pre-clear would otherwise + * cancel the auto-send for the new prompt. Transient — not persisted. + */ + dataLoadingChatPending: { text: string; images: string[]; attachments: string[] } | null; /** * Pending hand-off from the Data Agent to a peer agent. Set by the * Data Agent's `delegate` action card; consumed by `DataFormulator` @@ -299,6 +310,7 @@ const initialState: DataFormulatorState = { dataLoadingChatMessages: [], dataLoadingChatInProgress: false, dataLoadingChatResetCounter: 0, + dataLoadingChatPending: null, agentHandoffRequest: null, generatedReports: [], @@ -720,6 +732,7 @@ export const dataFormulatorSlice = createSlice({ state.dataLoadingChatMessages = []; state.dataLoadingChatInProgress = false; state.dataLoadingChatResetCounter = (state.dataLoadingChatResetCounter ?? 0) + 1; + state.dataLoadingChatPending = null; state.generatedReports = []; @@ -837,6 +850,7 @@ export const dataFormulatorSlice = createSlice({ config: { ...initialState.config, ...(saved.config || {}) }, dataCleanBlocks: saved.dataCleanBlocks || [], dataLoadingChatMessages: saved.dataLoadingChatMessages || [], + dataLoadingChatPending: null, generatedReports: saved.generatedReports || [], // Reset transient fields @@ -1665,6 +1679,20 @@ export const dataFormulatorSlice = createSlice({ state.dataLoadingChatMessages = []; state.dataLoadingChatInProgress = false; state.dataLoadingChatResetCounter = (state.dataLoadingChatResetCounter ?? 0) + 1; + // Note: `dataLoadingChatPending` is intentionally left + // alone. Callers that want "fresh slate + auto-send the + // new prompt" dispatch `clearChatMessages` followed by + // `setDataLoadingChatPending` in the same tick — clearing + // pending here would race with that ordering. + }, + setDataLoadingChatPending: ( + state, + action: PayloadAction<{ text: string; images: string[]; attachments: string[] }>, + ) => { + state.dataLoadingChatPending = action.payload; + }, + clearDataLoadingChatPending: (state) => { + state.dataLoadingChatPending = null; }, confirmTableLoad: (state, action: PayloadAction<{messageId: string, tableName: string}>) => { const msg = state.dataLoadingChatMessages.find(m => m.id === action.payload.messageId); diff --git a/src/views/DataFormulator.tsx b/src/views/DataFormulator.tsx index 2c21c15b..00e8086e 100644 --- a/src/views/DataFormulator.tsx +++ b/src/views/DataFormulator.tsx @@ -301,24 +301,37 @@ export const DataFormulatorFC = ({ }) => { // State for unified data upload dialog const [uploadDialogOpen, setUploadDialogOpen] = useState(false); const [uploadDialogInitialTab, setUploadDialogInitialTab] = useState('menu'); - const [uploadDialogInitialChatPrompt, setUploadDialogInitialChatPrompt] = useState(undefined); - const [uploadDialogInitialChatImages, setUploadDialogInitialChatImages] = useState(undefined); // Loading state for sessions (from Redux, shared with App.tsx) const sessionLoading = useSelector((state: DataFormulatorState) => state.sessionLoading); const sessionLoadingLabel = useSelector((state: DataFormulatorState) => state.sessionLoadingLabel); - const openUploadDialog = (tab: UploadTabType, initialChatPrompt?: string, initialChatImages?: string[]) => { + const openUploadDialog = (tab: UploadTabType) => { // If no workspace is active, generate an ID (backend creates folder lazily on first data op) if (!activeWorkspace) { dispatch(dfActions.setActiveWorkspace({ id: generateSessionId(), displayName: 'Untitled Session' })); } setUploadDialogInitialTab(tab); - setUploadDialogInitialChatPrompt(initialChatPrompt); - setUploadDialogInitialChatImages(initialChatImages); setUploadDialogOpen(true); }; + // Seed the Data Loading chat through the single redux `pending` slot, + // then navigate to the extract tab. This is the one channel that + // carries text, images, AND file attachments as first-class fields — + // replacing the older `initialChatPrompt/Images` props that silently + // dropped file attachments (they had no dedicated field and only + // survived if their name was baked into the prompt text). + const startDataLoadingChat = (text: string, images: string[] = [], attachments: string[] = []) => { + if (text.trim().length > 0 || images.length > 0 || attachments.length > 0) { + // Fresh query replaces any prior conversation. + if (dataLoadingChatMessages.length > 0) { + dispatch(dfActions.clearChatMessages()); + } + dispatch(dfActions.setDataLoadingChatPending({ text, images, attachments })); + } + openUploadDialog('extract'); + }; + // Honor cross-component requests to hand off to the Data Loading // chat seeded with a prompt (e.g. Data Agent's `delegate` card with // target='data_loading'). Hand-offs targeting other agents (e.g. @@ -326,7 +339,7 @@ export const DataFormulatorFC = ({ }) => { const agentHandoffRequest = useSelector((state: DataFormulatorState) => state.agentHandoffRequest); useEffect(() => { if (agentHandoffRequest && agentHandoffRequest.target === 'data_loading') { - openUploadDialog('extract', agentHandoffRequest.prompt, agentHandoffRequest.images); + startDataLoadingChat(agentHandoffRequest.prompt, agentHandoffRequest.images ?? [], []); dispatch(dfActions.clearAgentHandoffRequest()); } // openUploadDialog is stable enough for this purpose; we only react @@ -730,7 +743,7 @@ export const DataFormulatorFC = ({ }) => { openUploadDialog(`connector:${conn.id}` as UploadTabType); } }} - onStartChat={(prompt, images) => openUploadDialog('extract', prompt, images)} + onStartChat={(prompt, images, attachments) => startDataLoadingChat(prompt, images, attachments)} hasPriorConversation={dataLoadingChatMessages.length > 0} onResumeChat={() => openUploadDialog('extract')} serverConfig={serverConfig} @@ -933,16 +946,9 @@ export const DataFormulatorFC = ({ }) => { open={uploadDialogOpen} onClose={() => { setUploadDialogOpen(false); - // Clear one-shot seed values so the next dialog - // open (e.g. via the upload button) doesn't - // re-fire the agent with a stale prompt/image. - setUploadDialogInitialChatPrompt(undefined); - setUploadDialogInitialChatImages(undefined); refreshPageConnectors(); }} initialTab={uploadDialogInitialTab} - initialChatPrompt={uploadDialogInitialChatPrompt} - initialChatImages={uploadDialogInitialChatImages} onConnectorsChanged={handleConnectorsChanged} /> {/* Loading overlay for session loading */} diff --git a/src/views/DataLoadingChat.tsx b/src/views/DataLoadingChat.tsx index 379e29fb..9fe59000 100644 --- a/src/views/DataLoadingChat.tsx +++ b/src/views/DataLoadingChat.tsx @@ -60,7 +60,11 @@ const getUniqueTableName = (baseName: string, existingNames: Set): strin // Modern monospace font stack for code blocks const CODE_FONT = '"SF Mono", "Cascadia Code", "Fira Code", Menlo, Consolas, "Liberation Mono", monospace'; -const MarkdownContent: React.FC<{ content: string }> = ({ content }) => { +// Memoized so typing in the chat input (which re-renders the parent +// `DataLoadingChat` on every keystroke) doesn't re-parse every assistant +// message through react-markdown. `content` is a stable string per +// committed message, so the default shallow equality is sufficient. +const MarkdownContent = React.memo(({ content }: { content: string }) => { return ( = ({ content }) => { ); -}; +}); // --------------------------------------------------------------------------- // Inline table preview — compact notebook-style @@ -317,10 +321,16 @@ const CodeBlockView: React.FC<{ block: CodeExecution }> = ({ block }) => { // Single chat message bubble // --------------------------------------------------------------------------- -const ChatBubble: React.FC<{ +// Memoized so typing in the chat input doesn't re-render every prior +// bubble (each one renders MarkdownContent + potentially code blocks / +// table previews, which is expensive on long threads). The parent +// stabilises `existingNames` via useMemo so memo equality holds across +// keystrokes. +const ChatBubble = React.memo<{ message: ChatMessage; existingNames: Set; -}> = ({ message, existingNames }) => { + onTableLoaded?: () => void; +}>(({ message, existingNames, onTableLoaded }) => { const theme = useTheme(); const { t } = useTranslation(); const dispatch = useDispatch(); @@ -340,6 +350,9 @@ const ChatBubble: React.FC<{ if (table) { dispatch(loadTable({ table: { ...table, source: { type: 'extract' as const } } })); dispatch(dfActions.confirmTableLoad({ messageId: message.id, tableName: pending.name })); + // Loading data is a deliberate commit — return the + // user to the canvas (the dialog closes via this hook). + onTableLoaded?.(); } } } catch (err) { @@ -468,6 +481,11 @@ const ChatBubble: React.FC<{ })); } dispatch(dfActions.markLoadPlanConfirmed({ messageId: message.id })); + if (selected.length > 0) { + // Return the user to the canvas after a + // deliberate batch load. + onTableLoaded?.(); + } }} /> )} @@ -493,7 +511,7 @@ const ChatBubble: React.FC<{ ); -}; +}); // --------------------------------------------------------------------------- // Tool call label mapping @@ -517,7 +535,10 @@ interface ToolStep { label: string; } -const StreamingIndicator: React.FC<{ content: string; toolSteps: ToolStep[] }> = ({ content, toolSteps }) => { +// Memoized so an unrelated parent re-render (e.g. typing) doesn't +// reflow the shimmer animation. Props are state values that only change +// during an active stream. +const StreamingIndicator = React.memo<{ content: string; toolSteps: ToolStep[] }>(({ content, toolSteps }) => { const theme = useTheme(); return ( @@ -579,55 +600,56 @@ const StreamingIndicator: React.FC<{ content: string; toolSteps: ToolStep[] }> = )} ); -}; +}); // --------------------------------------------------------------------------- // Main chat component // --------------------------------------------------------------------------- -export interface DataLoadingChatProps { - /** - * Optional initial text to pre-fill the chat input when the component - * mounts (or when the value changes). Used by external entry points - * (e.g. landing page quick-chat box) that want to hand off a prompt - * to the agent. - */ - initialPrompt?: string; - /** - * Optional images (data URLs) to seed alongside `initialPrompt` — - * used when an external surface (e.g. landing-page agent box) has - * already collected pasted/attached images and is handing them off. - */ - initialImages?: string[]; - /** - * If true, automatically send the `initialPrompt` once on mount/change. - * Otherwise the prompt is only pre-filled and the user presses Enter. - */ - autoSendInitialPrompt?: boolean; +interface DataLoadingChatProps { + /** Called after a table is successfully loaded into the app. The + * upload dialog wires this to its close handler so loading data + * returns the user to the canvas. */ + onTableLoaded?: () => void; } -export const DataLoadingChat: React.FC = ({ - initialPrompt, - initialImages, - autoSendInitialPrompt, -}) => { +export const DataLoadingChat: React.FC = ({ onTableLoaded }) => { const theme = useTheme(); const { t } = useTranslation(); const dispatch = useDispatch(); + // Keep the latest callback in a ref so the stable `handleTableLoaded` + // identity below doesn't bust `ChatBubble`'s memoization even when the + // parent passes a fresh closure each render. + const onTableLoadedRef = useRef(onTableLoaded); + onTableLoadedRef.current = onTableLoaded; + const handleTableLoaded = useCallback(() => { + onTableLoadedRef.current?.(); + }, []); + const chatMessages = useSelector((state: DataFormulatorState) => state.dataLoadingChatMessages); const chatInProgress = useSelector((state: DataFormulatorState) => state.dataLoadingChatInProgress); - // External reset signal — bumped by `clearChatMessages` (manual reset - // button, new menu-level query, full session reset). When it changes - // we abort any in-flight stream, drop partial UI state, and re-seed - // from props if the parent provided a new prompt/images. Without - // this, an in-flight stream's eventual dispatches would leak into - // the freshly-cleared thread. + // External reset signal — bumped by `clearChatMessages` (manual + // reset button, fresh menu submission, full session reset). Used + // here only to abort an in-flight stream and invalidate any + // late-arriving dispatches from that stream via `sessionRef`. const chatResetCounter = useSelector((state: DataFormulatorState) => state.dataLoadingChatResetCounter ?? 0); + // Pending submission queued by an external surface (menu agent + // box, suggestion auto-run, external dialog caller). When set, we + // consume it in a useEffect: clear the slot first, then send the + // carried payload as a fresh user message via `sendMessage`. + // Single redux signal = no prop race. + const pendingSubmission = useSelector((state: DataFormulatorState) => state.dataLoadingChatPending); const existingTables = useSelector((state: DataFormulatorState) => state.tables); const activeModel = useSelector(dfSelectors.getActiveModel); const frontendRowLimit = useSelector((state: DataFormulatorState) => state.config?.frontendRowLimit ?? 2_000_000); - const existingNames = new Set(existingTables.map(tbl => tbl.id)); + // Stable reference across renders that don't actually change the + // table list — without this, every keystroke in the chat input + // would rebuild the Set and bust `ChatBubble`'s memo equality. + const existingNames = React.useMemo( + () => new Set(existingTables.map(tbl => tbl.id)), + [existingTables], + ); const [prompt, setPrompt] = useState(''); const [userImages, setUserImages] = useState([]); @@ -654,95 +676,44 @@ export const DataLoadingChat: React.FC = ({ // Auto-focus input useEffect(() => { inputRef.current?.focus(); }, []); - // ---- External initial prompt handling ------------------------------- - // Pre-fill the input (and optionally auto-send) when `initialPrompt` - // is provided. Used by external surfaces (e.g. landing-page quick chat - // box) to hand off text to the agent. Auto-send only fires for a - // fresh conversation — we never auto-resend on remount mid-chat. - const hasExistingMessages = chatMessages.length > 0; - const [pendingAutoSend, setPendingAutoSend] = useState(false); + // ---- Reset handling ------------------------------------------------- + // On external reset (counter bump from `clearChatMessages`): abort + // any in-flight stream, invalidate the current session token, and + // clear local input/streaming UI state. We deliberately do NOT + // re-seed anything here — a reset means "clean slate"; any new + // submission arrives separately via `pendingSubmission`. useEffect(() => { - // Detect external reset: abort, invalidate in-flight session, - // and clear all local UI state before re-seeding. Including - // `chatResetCounter` in the dep list also guarantees that an - // identical-prompt re-submission (same `initialPrompt` string) - // still triggers a fresh auto-send — otherwise the deps would - // be unchanged and the effect would skip. - const isReset = chatResetCounter !== lastResetRef.current; - if (isReset) { - lastResetRef.current = chatResetCounter; - sessionRef.current += 1; - abortControllerRef.current?.abort(); - abortControllerRef.current = null; - setStreamingContent(''); - setStreamingToolSteps([]); - setPrompt(''); - setUserImages([]); - setUserAttachments([]); - setPendingAutoSend(false); - } - - // Extract `[Uploaded: name]` mentions from the seeded prompt and - // surface them as chips. The mention template is locale-aware, - // so we build the regex from the current i18n value rather than - // hard-coding the English form. - const mentionTemplate = t('dataLoading.uploaded', { name: '__DF_NAME__' }); - const mentionPattern = mentionTemplate - .replace(/[.*+?^${}()|[\]\\]/g, '\\$&') - .replace('__DF_NAME__', '(.+?)'); - const mentionRegex = new RegExp(mentionPattern, 'g'); - let seededPrompt = initialPrompt || ''; - const extractedNames: string[] = []; - if (seededPrompt) { - let match: RegExpExecArray | null; - while ((match = mentionRegex.exec(seededPrompt)) !== null) { - extractedNames.push(match[1]); - } - if (extractedNames.length > 0) { - seededPrompt = seededPrompt - .replace(new RegExp(`\\n?${mentionPattern}`, 'g'), '') - .trim(); - } - } - - const hasText = seededPrompt.trim().length > 0; - const hasImages = !!initialImages && initialImages.length > 0; - const hasAttachments = extractedNames.length > 0; - // Skip re-seeding the input on a user-initiated reset — the - // reset is meant to restore a clean slate, not re-populate the - // input with the prompt the user just cleared. - if (!isReset) { - if (hasText) setPrompt(seededPrompt); - if (hasAttachments) setUserAttachments(extractedNames); - if (hasImages) { - // Always replace, never append. The prop is a "seed" — each - // change represents a fresh handoff from the parent, not an - // additive update. Appending caused the same image to stack - // up every time the parent re-rendered with a new array ref. - setUserImages([...initialImages!]); - } - } - // Auto-send only on a genuinely fresh open (no prior messages, - // and not a user-initiated reset). Resetting means the user wants - // a clean slate — re-running the seeded prompt against their will - // would defeat the purpose of the reset button. - if (autoSendInitialPrompt && !isReset && (hasText || hasImages || hasAttachments) && !hasExistingMessages) { - setPendingAutoSend(true); - } - // eslint-disable-next-line react-hooks/exhaustive-deps - }, [initialPrompt, initialImages, autoSendInitialPrompt, chatResetCounter]); + if (chatResetCounter === lastResetRef.current) return; + lastResetRef.current = chatResetCounter; + sessionRef.current += 1; + abortControllerRef.current?.abort(); + abortControllerRef.current = null; + setStreamingContent(''); + setStreamingToolSteps([]); + setPrompt(''); + setUserImages([]); + setUserAttachments([]); + }, [chatResetCounter]); const stopGeneration = () => { abortControllerRef.current?.abort(); }; // ---- Send message ---- - const sendMessage = useCallback(() => { - const text = prompt.trim(); - if (!text && userImages.length === 0 && userAttachments.length === 0) return; + // Accepts an optional explicit payload so callers (suggestion + // auto-run, pending-submission consume) can submit the exact + // values they just chose without waiting for React state to flush. + // Reading via the `prompt`/`userImages`/`userAttachments` closures + // alone would be racy with batching and could submit the previous + // round's values on a fresh handoff. + const sendMessage = useCallback((explicit?: { text: string; images: string[]; attachments: string[] }) => { + const text = (explicit?.text ?? prompt).trim(); + const imgs = explicit?.images ?? userImages; + const atts = explicit?.attachments ?? userAttachments; + if (!text && imgs.length === 0 && atts.length === 0) return; if (chatInProgress) return; - const imageAttachments: ChatAttachment[] = userImages.map((url, i) => ({ + const imageAttachments: ChatAttachment[] = imgs.map((url, i) => ({ type: 'image' as const, name: `image-${i + 1}`, url, })); - const fileAttachments: ChatAttachment[] = userAttachments.map(name => ({ + const fileAttachments: ChatAttachment[] = atts.map(name => ({ type: 'file' as const, name, })); const attachments: ChatAttachment[] = [...imageAttachments, ...fileAttachments]; @@ -751,7 +722,7 @@ export const DataLoadingChat: React.FC = ({ // chips (rendered from `attachments`). The agent payload below // re-injects `[Uploaded: name]` mentions so the backend still // sees the file references inline. - const displayText = text || (userImages.length > 0 ? t('dataLoading.defaultImageMessage') : ''); + const displayText = text || (imgs.length > 0 ? t('dataLoading.defaultImageMessage') : ''); const userMsg: ChatMessage = { id: `msg-${Date.now()}-user`, role: 'user', @@ -967,25 +938,48 @@ export const DataLoadingChat: React.FC = ({ } } })(); - }, [prompt, userImages, chatInProgress, chatMessages, activeModel, existingTables, dispatch, streamingContent, t]); + }, [prompt, userImages, userAttachments, chatInProgress, chatMessages, activeModel, existingTables, dispatch, streamingContent, t]); - // Auto-send the initial prompt once it has been applied to state. + // Consume a queued submission from any external surface (menu + // agent input, suggestion auto-run, or a cross-component handoff + // routed through `startDataLoadingChat`). Single redux signal, + // single consumer — no prop race. + // + // Idempotency note: under React.StrictMode (dev), effects are + // intentionally double-invoked on mount with the *same* closure, + // so the `clearDataLoadingChatPending` dispatch in the first run + // isn't visible to the second run. `lastConsumedRef` tracks the + // exact payload object we've already sent, so the second + // invocation short-circuits before calling `sendMessage` again. + const lastConsumedRef = useRef(null); useEffect(() => { - if (!pendingAutoSend) return; + if (!pendingSubmission) return; + if (pendingSubmission === lastConsumedRef.current) return; if (chatInProgress) return; - if (prompt.trim().length === 0 && userImages.length === 0) return; - setPendingAutoSend(false); - sendMessage(); - }, [pendingAutoSend, prompt, userImages, chatInProgress, sendMessage]); + lastConsumedRef.current = pendingSubmission; + const payload = pendingSubmission; + dispatch(dfActions.clearDataLoadingChatPending()); + sendMessage(payload); + }, [pendingSubmission, chatInProgress, sendMessage, dispatch]); // Reuse the shared sample-task list so this in-session panel stays in // sync with the upload-dialog entry point (`UnifiedDataUploadDialog`). + // Auto-run is wired through the redux pending slot so the click — + // even on a chat with prior history — atomically clears the thread + // and queues the new submission. const focusSuggestions = React.useMemo(() => buildDataLoadingSuggestions({ t, setInput: setPrompt, setImages: setUserImages, setAttachments: setUserAttachments, - }), [t]); + requestAutoSend: (payload) => { + if (chatMessages.length > 0) { + dispatch(dfActions.clearChatMessages()); + } + dispatch(dfActions.setDataLoadingChatPending(payload)); + }, + // eslint-disable-next-line react-hooks/exhaustive-deps + }), [t, dispatch]); const isEmpty = chatMessages.length === 0 && !streamingContent; @@ -1047,7 +1041,7 @@ export const DataLoadingChat: React.FC = ({ ) : ( <> {chatMessages.map((msg) => ( - + ))} {streamingContent !== '' && } {chatInProgress && !streamingContent && } @@ -1065,7 +1059,7 @@ export const DataLoadingChat: React.FC = ({ onChange={setPrompt} images={userImages} onImagesChange={setUserImages} - onSend={sendMessage} + onSend={() => sendMessage()} onStop={stopGeneration} inProgress={chatInProgress} placeholder={t('dataLoading.placeholder')} @@ -1076,8 +1070,13 @@ export const DataLoadingChat: React.FC = ({ formData.append('file', file); apiRequest(getUrls().SCRATCH_UPLOAD_URL, { method: 'POST', body: formData, - }).then(() => { - setUserAttachments(prev => [...prev, file.name]); + }).then(({ data }) => { + // The backend hash-suffixes the filename + // (e.g. `name_a1b2c3d4.xlsx`). Store the + // server-assigned name so the `[Uploaded:]` + // mention points to the real scratch file. + const scratchName = (data?.path || `scratch/${file.name}`).replace(/^scratch\//, ''); + setUserAttachments(prev => [...prev, scratchName]); }).catch(err => console.error('Upload failed:', err)); }} attachments={userAttachments} diff --git a/src/views/DataThread.tsx b/src/views/DataThread.tsx index ae0cc9f1..c69a71a5 100644 --- a/src/views/DataThread.tsx +++ b/src/views/DataThread.tsx @@ -1751,6 +1751,9 @@ let SingleThreadGroupView: FC<{ const TIMELINE_GAP = '4px'; // gap between timeline and card content const DOT_SIZE = 6; const CARD_PY = '6px'; // vertical padding for each timeline row + // Mirror the left timeline gutter on the right so cards sit visually + // centred in their column instead of hugging the right edge. + const CARD_CONTENT_PR = `${TIMELINE_WIDTH}px`; // CSS `border-style: dashed` stretches dashes to fit each element's // height, so stacked segments end up with mismatched dash lengths. A @@ -1907,7 +1910,7 @@ let SingleThreadGroupView: FC<{ {isLast && hasContinuationBelow && } {isLast && !hasContinuationBelow && } - + {item.element} @@ -1983,7 +1986,7 @@ let SingleThreadGroupView: FC<{ {isLast && hasContinuationBelow && } {isLast && !hasContinuationBelow && } - + {item.element} @@ -2006,7 +2009,7 @@ let SingleThreadGroupView: FC<{ {isLast && hasContinuationBelow && } {isLast && !hasContinuationBelow && } - + {item.element} @@ -2054,7 +2057,7 @@ let SingleThreadGroupView: FC<{ )} {isLast && !hasContinuationBelow && } - {item.element} @@ -3119,7 +3122,10 @@ export const DataThread: FC<{sx?: SxProps}> = function ({ sx }) { // benefit, since the segments would just stack in the same single column. const CARD_GAP = 12; // padding + spacing between cards in a column const PANEL_PADDING = 16; - const CARD_WIDTH = 220; + // 220 visual card width + 14px right gutter (CARD_CONTENT_PR) so cards + // keep their original size while gaining a right margin that balances + // the left timeline gutter. + const CARD_WIDTH = 234; const COLUMN_WIDTH = CARD_WIDTH + CARD_GAP; // n columns need: n*CARD_WIDTH + (n-1)*CARD_GAP + PANEL_PADDING // Solving for n: n <= (containerWidth - PANEL_PADDING + CARD_GAP) / COLUMN_WIDTH diff --git a/src/views/UnifiedDataUploadDialog.tsx b/src/views/UnifiedDataUploadDialog.tsx index bd7167f8..73d325e1 100644 --- a/src/views/UnifiedDataUploadDialog.tsx +++ b/src/views/UnifiedDataUploadDialog.tsx @@ -448,12 +448,14 @@ export interface DataLoadMenuProps { onSelectConnector?: (connector: ConnectorInstance) => void; /** * Called when the user submits a prompt from the top-level Data Loading - * Agent chat box. Implementations should open the agent chat surface - * with the prompt (and optional pasted/attached images) pre-filled — - * typically auto-sent. If not provided, the chat box falls back to - * `onSelectTab('extract')`. + * Agent chat box. Implementations should hand the payload off to the + * agent chat surface, which will auto-send it as a fresh user + * message. Attachments are file names (already uploaded to the + * session scratch space) — the chat surface re-injects them as + * `[Uploaded: name]` mentions when building the backend payload. + * If not provided, the chat box falls back to `onSelectTab('extract')`. */ - onStartChat?: (prompt: string, images?: string[]) => void; + onStartChat?: (prompt: string, images: string[], attachments: string[]) => void; /** * True when a prior data-loading agent conversation exists in * state. When set together with `onResumeChat`, the menu renders @@ -605,22 +607,17 @@ export const DataLoadMenu: React.FC = ({ const submitAgentChat = () => { const text = agentInput.trim(); if (text.length === 0 && agentImages.length === 0 && agentAttachments.length === 0) { - // Empty submission — just open the chat surface. - if (onStartChat) onStartChat('', []); + // Empty submission — just surface the chat. + if (onStartChat) onStartChat('', [], []); else onSelectTab('extract'); return; } - // Augment the outgoing prompt with `[Uploaded: name]` lines so the - // agent sees attachments as text references, without polluting - // the editable input the user sees. - const mentions = agentAttachments - .map(name => t('dataLoading.uploaded', { name })) - .join('\n'); - const finalText = mentions - ? (text ? `${text}\n${mentions}` : mentions) - : text; + // Pass payload pieces unchanged — the chat surface builds the + // backend mentions itself. We deliberately do NOT pre-inject + // `[Uploaded: name]` into `text` here, so the visible message + // bubble stays clean and the file chips render uniformly. if (onStartChat) { - onStartChat(finalText, agentImages); + onStartChat(text, agentImages, agentAttachments); } else { onSelectTab('extract'); } @@ -631,14 +628,26 @@ export const DataLoadMenu: React.FC = ({ // Suggestions surfaced as a focus-time dropdown — sourced from a shared // factory so the in-session `DataLoadingChat` panel renders the exact - // same list. See `dataLoadingSuggestions.ts`. + // same list. See `dataLoadingSuggestions.ts`. Auto-run is routed + // through `onStartChat` so the parent dialog can dispatch its + // `clearChatMessages` + `setDataLoadingChatPending` sequence + // atomically — same path as a manual submit. const agentChatSuggestions = useMemo(() => buildDataLoadingSuggestions({ t, setInput: setAgentInput, setImages: setAgentImages, setAttachments: setAgentAttachments, ensureActiveWorkspace, - }), [t]); + requestAutoSend: onStartChat + ? (payload) => { + onStartChat(payload.text, payload.images, payload.attachments); + setAgentInput(''); + setAgentImages([]); + setAgentAttachments([]); + } + : undefined, + // eslint-disable-next-line react-hooks/exhaustive-deps + }), [t, onStartChat]); const agentChatBox = ( = ({ formData.append('file', file); apiRequest(getUrls().SCRATCH_UPLOAD_URL, { method: 'POST', body: formData, - }).then(() => { - setAgentAttachments(prev => [...prev, file.name]); + }).then(({ data }) => { + // The backend hash-suffixes the filename; store the + // server-assigned name so the `[Uploaded:]` mention + // resolves to the real scratch file. + const scratchName = (data?.path || `scratch/${file.name}`).replace(/^scratch\//, ''); + setAgentAttachments(prev => [...prev, scratchName]); }).catch(err => console.error('Upload failed:', err)); }} attachments={agentAttachments} @@ -1112,14 +1125,6 @@ export interface UnifiedDataUploadDialogProps { open: boolean; onClose: () => void; initialTab?: UploadTabType; - /** - * Optional initial prompt to hand off to the Data Loading Agent. When - * non-empty and `initialTab === 'extract'`, the prompt is pre-filled - * and auto-sent in the chat panel. - */ - initialChatPrompt?: string; - /** Optional images (data URLs) to seed the chat alongside `initialChatPrompt`. */ - initialChatImages?: string[]; onConnectorsChanged?: () => void; } @@ -1127,8 +1132,6 @@ export const UnifiedDataUploadDialog: React.FC = ( open, onClose, initialTab = 'menu', - initialChatPrompt, - initialChatImages, onConnectorsChanged, }) => { const theme = useTheme(); @@ -1143,21 +1146,6 @@ export const UnifiedDataUploadDialog: React.FC = ( const existingNames = new Set(existingTables.map(t => t.id)); const [activeTab, setActiveTab] = useState(initialTab === 'menu' ? 'menu' : initialTab); - // Prompt to seed the agent chat with. Sourced from the `initialChatPrompt` - // prop when the dialog opens directly on 'extract', or set internally - // when the user submits the in-menu agent chat box. - const [seededChatPrompt, setSeededChatPrompt] = useState( - initialTab === 'extract' ? initialChatPrompt : undefined, - ); - const [seededChatImages, setSeededChatImages] = useState( - initialTab === 'extract' ? initialChatImages : undefined, - ); - const [autoSendSeededPrompt, setAutoSendSeededPrompt] = useState( - initialTab === 'extract' && ( - (!!initialChatPrompt && initialChatPrompt.trim().length > 0) - || (!!initialChatImages && initialChatImages.length > 0) - ), - ); const fileInputRef = useRef(null); const urlInputRef = useRef(null); @@ -1175,27 +1163,8 @@ export const UnifiedDataUploadDialog: React.FC = ( if (open) { setConnectorInstances([]); refreshConnectors(); - // Re-seed chat prompt/images from props each time the dialog opens. - if (initialTab === 'extract') { - setSeededChatPrompt(initialChatPrompt); - setSeededChatImages(initialChatImages); - const hasText = !!initialChatPrompt && initialChatPrompt.trim().length > 0; - const hasImages = !!initialChatImages && initialChatImages.length > 0; - setAutoSendSeededPrompt(hasText || hasImages); - // Opening the dialog with a fresh prompt/images means the - // user wants a new data-loading conversation; clear any - // stale messages from a previous session so the new query - // isn't appended to an unrelated thread. - if ((hasText || hasImages) && dataLoadingChatMessages.length > 0) { - dispatch(dfActions.clearChatMessages()); - } - } else { - setSeededChatPrompt(undefined); - setSeededChatImages(undefined); - setAutoSendSeededPrompt(false); - } } - }, [open, refreshConnectors, identityKey, initialTab, initialChatPrompt, initialChatImages]); + }, [open, refreshConnectors, identityKey]); // Storage is determined by backend config — no user toggle const isEphemeral = serverConfig.WORKSPACE_BACKEND === 'ephemeral'; @@ -1848,29 +1817,32 @@ export const UnifiedDataUploadDialog: React.FC = ( setActiveTab(`connector:${conn.id}` as UploadTabType); } }} - onStartChat={(prompt, images) => { + onStartChat={(prompt, images, attachments) => { const hasText = prompt.trim().length > 0; - const hasImages = !!images && images.length > 0; - // If a prior conversation exists, treat a - // new query from the menu as a fresh data - // reload and reset the chat. Without this - // the new prompt would be appended onto an - // unrelated thread, confusing the agent. - if ((hasText || hasImages) && dataLoadingChatMessages.length > 0) { - dispatch(dfActions.clearChatMessages()); + const hasImages = images.length > 0; + const hasAttachments = attachments.length > 0; + // Always surface the chat. If the user + // is starting a fresh query, clear any + // prior conversation and enqueue the new + // submission as a redux `pending` slot + // — `DataLoadingChat` consumes it on + // render and auto-sends. Doing both + // dispatches in the same tick keeps the + // handoff atomic; there's no prop race. + if (hasText || hasImages || hasAttachments) { + if (dataLoadingChatMessages.length > 0) { + dispatch(dfActions.clearChatMessages()); + } + dispatch(dfActions.setDataLoadingChatPending({ + text: prompt, images, attachments, + })); } - setSeededChatPrompt(prompt); - setSeededChatImages(images); - setAutoSendSeededPrompt(hasText || hasImages); setActiveTab('extract'); }} hasPriorConversation={dataLoadingChatMessages.length > 0} onResumeChat={() => { // Reopen the existing thread without // clearing messages or auto-sending. - setSeededChatPrompt(undefined); - setSeededChatImages(undefined); - setAutoSendSeededPrompt(false); setActiveTab('extract'); }} serverConfig={serverConfig} @@ -2403,11 +2375,7 @@ export const UnifiedDataUploadDialog: React.FC = ( {/* Extract Data Tab */} - + {/* Local Folder Tab */} diff --git a/src/views/dataLoadingSuggestions.ts b/src/views/dataLoadingSuggestions.ts index 8d91b92b..f37e04f0 100644 --- a/src/views/dataLoadingSuggestions.ts +++ b/src/views/dataLoadingSuggestions.ts @@ -22,6 +22,12 @@ export interface DataLoadingSuggestion { onClick: () => void; } +export interface SuggestionPayload { + text: string; + images: string[]; + attachments: string[]; +} + export interface BuildSuggestionsArgs { t: TFunction; setInput: (value: string) => void; @@ -29,12 +35,22 @@ export interface BuildSuggestionsArgs { setAttachments: (names: string[]) => void; /** Optional hook that workspaces use to make sure a session exists before uploading. */ ensureActiveWorkspace?: () => void; + /** + * Optional auto-run hook. When provided, suggestions submit the + * complete payload immediately (after any required async upload / + * data-URL prep) instead of just pre-filling the input. Callers + * typically wire this to a redux pending-submission dispatch so the + * payload survives the parent→child handoff without prop races. + * When absent, the suggestion behaves like a paste: it only fills + * the input fields via the `set*` callbacks. + */ + requestAutoSend?: (payload: SuggestionPayload) => void; } const EXCEL_SAMPLE_NAME = 'climate-gas-indicator.xlsx'; export function buildDataLoadingSuggestions( - { t, setInput, setImages, setAttachments, ensureActiveWorkspace }: BuildSuggestionsArgs, + { t, setInput, setImages, setAttachments, ensureActiveWorkspace, requestAutoSend }: BuildSuggestionsArgs, ): DataLoadingSuggestion[] { const kindAsk = t('upload.agentChatSuggestion.kind.ask', { defaultValue: 'ask' }); const kindFind = t('upload.agentChatSuggestion.kind.find', { defaultValue: 'find' }); @@ -61,37 +77,38 @@ export function buildDataLoadingSuggestions( const iconSx = { fontSize: 14 }; + // Common: fill the input fields AND (if auto-run is enabled) submit + // the payload. Centralising the dual behaviour keeps every + // suggestion below short and consistent. + const fillAndMaybeSend = (payload: SuggestionPayload) => { + setImages(payload.images); + setAttachments(payload.attachments); + setInput(payload.text); + requestAutoSend?.(payload); + }; + return [ { kind: kindAsk, label: askLabel, icon: React.createElement(QuestionAnswerOutlinedIcon, { sx: iconSx }), - onClick: () => { - setImages([]); - setAttachments([]); - setInput(askLabel); - }, + onClick: () => fillAndMaybeSend({ text: askLabel, images: [], attachments: [] }), }, { kind: kindFind, label: findLabel, icon: React.createElement(SearchIcon, { sx: iconSx }), - onClick: () => { - setImages([]); - setAttachments([]); - setInput(findLabel); - }, + onClick: () => fillAndMaybeSend({ text: findLabel, images: [], attachments: [] }), }, { kind: kindExtract, label: extractExcelLabel, icon: React.createElement(TableChartOutlinedIcon, { sx: iconSx }), onClick: () => { - // Surface the attachment chip synchronously so it is - // always present when the user hits send, even if the - // upload below is still mid-flight. The chip is what - // gets serialised into the outgoing `[Uploaded: name]` - // mention and ultimately the chat bubble. + // Surface the attachment chip / input synchronously so + // it is visible during the async upload. The auto-send + // (if enabled) waits until the upload completes so the + // backend can actually find the scratch file. setImages([]); setAttachments([EXCEL_SAMPLE_NAME]); setInput(extractExcelLabel); @@ -108,6 +125,18 @@ export function buildDataLoadingSuggestions( method: 'POST', body: formData, }); }) + .then(({ data }) => { + // The backend hash-suffixes the filename, so use the + // server-assigned name for the chip and the mention + // — otherwise the agent looks for a file that the + // upload renamed and reports it missing. + const scratchName = (data?.path || `scratch/${EXCEL_SAMPLE_NAME}`).replace(/^scratch\//, ''); + setAttachments([scratchName]); + requestAutoSend?.({ + text: extractExcelLabel, images: [], + attachments: [scratchName], + }); + }) .catch(err => console.error('Sample Excel upload failed:', err)); }, }, @@ -116,16 +145,21 @@ export function buildDataLoadingSuggestions( label: extractImageLabel, icon: React.createElement(ImageOutlinedIcon, { sx: iconSx }), onClick: () => { + // Image needs to be read into a data URL before we can + // surface it as a chip or send it. Defer auto-send until + // the FileReader resolves. fetch(exampleImageTable) .then(res => res.blob()) .then(blob => { const reader = new FileReader(); reader.onload = () => { - if (reader.result) { - setImages([reader.result as string]); - setAttachments([]); - setInput(extractImageLabel); - } + if (!reader.result) return; + const dataUrl = reader.result as string; + fillAndMaybeSend({ + text: extractImageLabel, + images: [dataUrl], + attachments: [], + }); }; reader.readAsDataURL(blob); }); @@ -135,11 +169,7 @@ export function buildDataLoadingSuggestions( kind: kindExtract, label: extractTextLabel, icon: React.createElement(DescriptionOutlinedIcon, { sx: iconSx }), - onClick: () => { - setImages([]); - setAttachments([]); - setInput(extractTextPrompt); - }, + onClick: () => fillAndMaybeSend({ text: extractTextPrompt, images: [], attachments: [] }), }, ]; } diff --git a/tests/backend/data/test_workspace_manager.py b/tests/backend/data/test_workspace_manager.py index e2a00eb7..d82766c6 100644 --- a/tests/backend/data/test_workspace_manager.py +++ b/tests/backend/data/test_workspace_manager.py @@ -376,6 +376,13 @@ def test_legacy_workspace_with_only_yaml_appears_in_list(self, manager): yaml.safe_dump({"version": "1.1", "tables": {}}), encoding="utf-8", ) + # Pretend the legacy workspace had session state with tables. + (ws_dir / "session_state.json").write_text( + json.dumps({"tables": [{"id": "t1"}]}), + encoding="utf-8", + ) + # Trigger meta repair with a non-empty table count. + manager.save_session_state("legacy_ws", {"tables": [{"id": "t1"}]}) ws_list = manager.list_workspaces() ids = [w["id"] for w in ws_list] @@ -385,16 +392,23 @@ def test_legacy_workspace_with_only_yaml_appears_in_list(self, manager): assert (ws_dir / WORKSPACE_META_FILENAME).exists() def test_legacy_workspace_with_only_session_state_appears_in_list(self, manager): - """A directory with only session_state.json should be auto-repaired.""" + """A directory with session_state.json (containing tables) is + auto-repaired and visible in list_workspaces. The displayName + is inferred from session_state.""" ws_dir = manager.root / "state_only" ws_dir.mkdir(parents=True) (ws_dir / "session_state.json").write_text( json.dumps({ - "tables": [], + "tables": [{"id": "t1", "name": "T1"}], "activeWorkspace": {"displayName": "My Old Session"}, }), encoding="utf-8", ) + # Re-save so meta is written with tableCount > 0. + manager.save_session_state("state_only", { + "tables": [{"id": "t1", "name": "T1"}], + "activeWorkspace": {"displayName": "My Old Session"}, + }) ws_list = manager.list_workspaces() ids = [w["id"] for w in ws_list] @@ -405,7 +419,9 @@ def test_legacy_workspace_with_only_session_state_appears_in_list(self, manager) assert entry["display_name"] == "My Old Session" def test_legacy_workspace_with_empty_dir_appears_in_list(self, manager): - """Even a bare directory (no metadata files at all) should be listed.""" + """A bare directory with no metadata at all is auto-repaired by + _ensure_meta (meta.json gets created with fallback displayName) + and appears in list_workspaces.""" ws_dir = manager.root / "bare" ws_dir.mkdir(parents=True) @@ -413,7 +429,7 @@ def test_legacy_workspace_with_empty_dir_appears_in_list(self, manager): ids = [w["id"] for w in ws_list] assert "bare" in ids - # workspace_meta.json auto-created with fallback displayName = dir name + # Auto-repair created the meta with a fallback displayName. meta = json.loads((ws_dir / WORKSPACE_META_FILENAME).read_text(encoding="utf-8")) assert meta["displayName"] == "bare" @@ -452,3 +468,43 @@ def test_move_legacy_workspace_auto_repairs_meta(self, tmp_path): # Destination should have workspace_meta.json dst_ws = dst.get_workspace_path("old_ws") assert (dst_ws / WORKSPACE_META_FILENAME).exists() + + +class TestEmptyWorkspaceVisibility: + """list_workspaces() lists every workspace directory, including + empty "Untitled Session" entries from abandoned data-loading + chats. Users manage (rename/delete) these themselves via the + sidebar — they are not hidden.""" + + def test_empty_workspace_is_visible(self, manager): + manager.create_workspace("ghost") + # No save_session_state — meta has no tableCount/chartCount. + + ws_list = manager.list_workspaces() + + assert any(w["id"] == "ghost" for w in ws_list) + assert manager.workspace_exists("ghost") + + def test_workspace_with_tables_is_visible(self, manager): + manager.create_workspace("real") + manager.save_session_state("real", { + "tables": [{"id": "t1", "name": "T1"}], + "activeWorkspace": {"id": "real", "displayName": "Real"}, + }) + + ws_list = manager.list_workspaces() + + assert any(w["id"] == "real" for w in ws_list) + + def test_zero_count_workspace_is_visible(self, manager): + """A workspace whose tables were all deleted (zero tables) still + appears in the list — the user decides whether to remove it.""" + manager.create_workspace("emptied") + manager.save_session_state("emptied", { + "tables": [], + "activeWorkspace": {"id": "emptied", "displayName": "Emptied"}, + }) + + ws_list = manager.list_workspaces() + + assert any(w["id"] == "emptied" for w in ws_list) From 18cf3603dfd7da2197f4ba131e80ecb2fab285f3 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Thu, 28 May 2026 22:21:14 -0700 Subject: [PATCH 02/29] small fixes --- src/views/DataFormulator.tsx | 4 +-- src/views/DataSourceSidebar.tsx | 64 ++++++++------------------------- 2 files changed, 17 insertions(+), 51 deletions(-) diff --git a/src/views/DataFormulator.tsx b/src/views/DataFormulator.tsx index 00e8086e..b340a477 100644 --- a/src/views/DataFormulator.tsx +++ b/src/views/DataFormulator.tsx @@ -571,7 +571,7 @@ export const DataFormulatorFC = ({ }) => { const fixedSplitPane = ( openUploadDialog((tab ?? 'add-connection') as UploadTabType)} + onOpenUploadDialog={(tab) => openUploadDialog((tab ?? 'menu') as UploadTabType)} connectorRefreshKey={connectorRefreshKey} /> { {tables.length > 0 ? fixedSplitPane : ( openUploadDialog((tab ?? 'add-connection') as UploadTabType)} + onOpenUploadDialog={(tab) => openUploadDialog((tab ?? 'menu') as UploadTabType)} connectorRefreshKey={connectorRefreshKey} /> {dataUploadRequestBox} diff --git a/src/views/DataSourceSidebar.tsx b/src/views/DataSourceSidebar.tsx index a8c4715d..7381a423 100644 --- a/src/views/DataSourceSidebar.tsx +++ b/src/views/DataSourceSidebar.tsx @@ -42,7 +42,6 @@ import { VirtualizedCatalogTree } from '../components/VirtualizedCatalogTree'; import StorageIcon from '@mui/icons-material/Storage'; import AddIcon from '@mui/icons-material/Add'; -import FileUploadOutlinedIcon from '@mui/icons-material/FileUploadOutlined'; import FolderOpenIcon from '@mui/icons-material/FolderOpen'; import FolderOutlinedIcon from '@mui/icons-material/FolderOutlined'; import UploadFileIcon from '@mui/icons-material/UploadFile'; @@ -51,9 +50,6 @@ import ChevronLeftIcon from '@mui/icons-material/ChevronLeft'; import ExpandMoreIcon from '@mui/icons-material/ExpandMore'; import ChevronRightIcon from '@mui/icons-material/ChevronRight'; import RefreshIcon from '@mui/icons-material/Refresh'; -import ContentPasteOutlinedIcon from '@mui/icons-material/ContentPasteOutlined'; -import SmartToyOutlinedIcon from '@mui/icons-material/SmartToyOutlined'; -import LinkOutlinedIcon from '@mui/icons-material/LinkOutlined'; import LinkOffOutlinedIcon from '@mui/icons-material/LinkOffOutlined'; import DeleteOutlineIcon from '@mui/icons-material/DeleteOutline'; import EditOutlinedIcon from '@mui/icons-material/EditOutlined'; @@ -159,7 +155,7 @@ export const DataSourceSidebar: React.FC<{ // built-in sample_datasets connector is shown there, giving users // something useful to explore immediately. The upgrade message only // appears when they try to add a new connector or link a folder. - const [initialTab, setInitialTab] = useState<'upload' | 'sources' | 'sessions' | 'knowledge'>('sources'); + const [initialTab, setInitialTab] = useState<'sources' | 'sessions' | 'knowledge'>('sources'); // External callers (e.g. SaveExperienceButton on success) can ask the // sidebar to open and switch to a specific tab. @@ -277,6 +273,18 @@ export const DataSourceSidebar: React.FC<{ pt: 1, gap: 0.5, }}> + {/* Primary action — adding data is the main task. Styled like + the view-switcher icons but kept in primary color as a + subtle cue; opens the upload dialog (landing menu). */} + + onOpenUploadDialog?.()} sx={{ + color: 'primary.main', + borderRadius: 1, + '&:hover': { bgcolor: 'action.hover' }, + }}> + + + { setInitialTab('sessions'); if (!isOpen) toggle(); else if (initialTab !== 'sessions') setInitialTab('sessions'); else toggle(); }} sx={{ color: isOpen && initialTab === 'sessions' ? 'primary.main' : 'text.secondary', @@ -295,15 +303,6 @@ export const DataSourceSidebar: React.FC<{ - - { setInitialTab('upload'); if (!isOpen) toggle(); else if (initialTab !== 'upload') setInitialTab('upload'); else toggle(); }} sx={{ - color: isOpen && initialTab === 'upload' ? 'primary.main' : 'text.secondary', - bgcolor: isOpen && initialTab === 'upload' ? 'action.selected' : 'transparent', - borderRadius: 1, - }}> - - - { setInitialTab('knowledge'); if (!isOpen) toggle(); else if (initialTab !== 'knowledge') setInitialTab('knowledge'); else toggle(); }} sx={{ color: isOpen && initialTab === 'knowledge' ? 'primary.main' : 'text.secondary', @@ -347,7 +346,7 @@ const DataSourceSidebarPanel: React.FC<{ panelWidth: number; onOpenUploadDialog?: (tab?: string) => void; onCollapse: () => void; - initialTab?: 'upload' | 'sources' | 'sessions' | 'knowledge'; + initialTab?: 'sources' | 'sessions' | 'knowledge'; connectorRefreshKey?: number; disableConnectors?: boolean; }> = ({ panelWidth, onOpenUploadDialog, onCollapse, initialTab = 'sources', connectorRefreshKey = 0, disableConnectors = false }) => { @@ -419,7 +418,7 @@ const DataSourceSidebarPanel: React.FC<{ const [searchingCatalog, setSearchingCatalog] = useState>({}); // Sidebar tab: 'sources' or 'sessions' or 'knowledge' - const [activeTab, setActiveTab] = useState<'upload' | 'sources' | 'sessions' | 'knowledge'>(initialTab); + const [activeTab, setActiveTab] = useState<'sources' | 'sessions' | 'knowledge'>(initialTab); // Sync tab when rail icon switches it useEffect(() => { @@ -1292,39 +1291,6 @@ const DataSourceSidebarPanel: React.FC<{ overflow: 'hidden', }}> - {/* ── Upload Data tab ── */} - {activeTab === 'upload' && ( - - - - {t('sidebar.uploadData', { defaultValue: 'Upload Data' })} - - - - - - - - - {[ - { icon: , label: t('upload.uploadFile', { defaultValue: 'Upload file' }), tab: 'upload' }, - { icon: , label: t('upload.pasteData', { defaultValue: 'Paste data' }), tab: 'paste' }, - { icon: , label: t('upload.extractData', { defaultValue: 'Data Assistant' }), tab: 'extract' }, - { icon: , label: t('upload.loadFromUrl', { defaultValue: 'Load from URL' }), tab: 'url' }, - ].map((item, i) => ( - onOpenUploadDialog?.(item.tab)} - sx={{ display: 'flex', alignItems: 'center', gap: 0.75, px: 1.5, py: 0.75, cursor: 'pointer', color: 'text.primary', '&:hover': { bgcolor: 'action.hover' }, userSelect: 'none' }} - > - {item.icon} - {item.label} - - ))} - - - )} - {/* ── Data Connectors tab ── Sample datasets remain available even when external connectors are disabled; the Add Connector / Link Folder From 4cb0a2f4f5e32bd5132a84b77cdcb3cc12f50f56 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Thu, 28 May 2026 22:50:07 -0700 Subject: [PATCH 03/29] cleanup --- src/views/ChartRecBox.tsx | 14 +++++++------- src/views/SimpleChartRecBox.tsx | 34 +++++---------------------------- 2 files changed, 12 insertions(+), 36 deletions(-) diff --git a/src/views/ChartRecBox.tsx b/src/views/ChartRecBox.tsx index 872dc3c1..9eb26085 100644 --- a/src/views/ChartRecBox.tsx +++ b/src/views/ChartRecBox.tsx @@ -292,10 +292,10 @@ export const ChartRecBox: FC = function ({ tableId, placeHolde type={current ? undefined : 'button'} onClick={current ? undefined : () => dispatch(dfActions.setFocused({ type: 'table', tableId: table.id }))} sx={{ - display: 'inline-flex', alignItems: 'center', gap: current ? '6px' : '3px', + display: 'inline-flex', alignItems: 'center', gap: '3px', border: 'none', background: 'transparent', p: 0, fontFamily: theme.typography.fontFamily, - fontSize: current ? 16 : 11, lineHeight: 1.4, + fontSize: 11, lineHeight: 1.4, color: current ? 'primary.main' : 'text.secondary', fontWeight: current ? 600 : 400, cursor: current ? 'default' : 'pointer', @@ -304,7 +304,7 @@ export const ChartRecBox: FC = function ({ tableId, placeHolde '&:hover': current ? undefined : { color: 'primary.main' }, }} > - + {table.displayId} ); @@ -682,10 +682,10 @@ export const ChartRecBox: FC = function ({ tableId, placeHolde ); }; - // Center cluster auto-scales with chart count; neighbour - // clusters are halved and dimmed to read as context. - const centerN = Math.min(chartsForTable(currentTable.id).length, 8); - const centerScale = centerN <= 3 ? 1 : centerN <= 5 ? 0.82 : 0.66; + // All clusters render at the same scale; the current + // cluster is only distinguished by not being dimmed and by + // showing more thumbnails. + const centerScale = 0.5; const sideScale = 0.5; return ( diff --git a/src/views/SimpleChartRecBox.tsx b/src/views/SimpleChartRecBox.tsx index c96dec65..5e7bd63b 100644 --- a/src/views/SimpleChartRecBox.tsx +++ b/src/views/SimpleChartRecBox.tsx @@ -41,7 +41,7 @@ import StopIcon from '@mui/icons-material/Stop'; import AutoGraphIcon from '@mui/icons-material/AutoGraph'; import DescriptionOutlinedIcon from '@mui/icons-material/DescriptionOutlined'; import { UnifiedDataUploadDialog } from './UnifiedDataUploadDialog'; -import { transition } from '../app/tokens'; +import { borderColor, transition } from '../app/tokens'; import { Theme } from '@mui/material/styles'; import { useTranslation } from 'react-i18next'; import { shouldAutoFocusGeneratedChart } from '../app/agentInteractionPolicy'; @@ -1380,12 +1380,6 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ }, [pendingClarification, dispatch, t]); const isReportMode = selectedAgent === 'report'; - const gradientBorder = isReportMode - ? `linear-gradient(135deg, ${alpha(theme.palette.warning.main, 0.6)}, ${alpha(theme.palette.warning.dark, 0.5)})` - : `linear-gradient(135deg, ${alpha(theme.palette.primary.main, 0.6)}, ${alpha(theme.palette.secondary.main, 0.55)})`; - const workingBorder = isReportMode - ? `linear-gradient(135deg, ${alpha(theme.palette.warning.main, 0.3)}, ${alpha(theme.palette.warning.dark, 0.25)})` - : `linear-gradient(135deg, ${alpha(theme.palette.primary.main, 0.3)}, ${alpha(theme.palette.secondary.main, 0.25)})`; // Landing / "no thread yet" highlight: when the user has loaded data // but hasn't started an exploration on the focused table (no real @@ -1419,12 +1413,9 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ mx: 1, mb: 1, mt: 0.5, px: 1.25, pt: 1, pb: 0.5, borderRadius: '12px', - // The 2-tone border is drawn by the `::before` gradient - // overlay below (works through border-radius + masks). We - // intentionally leave the Card's own border off so the two - // don't fight; focus state uses a shadow halo instead of a - // border-color shift. - border: 'none', + // Standard single-tone input style (matches AgentChatInput): a + // solid divider border that turns the accent color on focus. + border: `1px solid ${borderColor.divider}`, outline: 'none', position: 'relative', overflow: isChatFormulating ? 'hidden' : 'visible', @@ -1454,24 +1445,9 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ } : {}), '&:focus-within': { animation: 'none', + borderColor: isReportMode ? theme.palette.warning.main : theme.palette.primary.main, boxShadow: `0 0 0 2px ${alpha(isReportMode ? theme.palette.warning.main : theme.palette.primary.main, 0.15)}, 0 2px 10px rgba(32, 33, 36, 0.14)`, }, - // Gradient border via pseudo-element (works with border-radius) - '&::before': { - content: '""', - position: 'absolute', - inset: 0, - borderRadius: 'inherit', - padding: '1.5px', - background: isChatFormulating - ? workingBorder - : gradientBorder, - WebkitMask: 'linear-gradient(#fff 0 0) content-box, linear-gradient(#fff 0 0)', - WebkitMaskComposite: 'xor', - maskComposite: 'exclude', - pointerEvents: 'none', - zIndex: 3, - }, }} > {clarificationQuestions?.kind === 'clarification' && clarificationQuestions.questions && pendingClarification && !isChatFormulating && ( From c616338d67c4c12cb7290139ce046213abc44098 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Fri, 29 May 2026 00:06:50 -0700 Subject: [PATCH 04/29] some updates --- src/views/DataView.tsx | 92 +++++++++++++++++++++++++++++++-- src/views/VisualizationView.tsx | 20 ++++--- 2 files changed, 98 insertions(+), 14 deletions(-) diff --git a/src/views/DataView.tsx b/src/views/DataView.tsx index aa1263d0..f6c4a79f 100644 --- a/src/views/DataView.tsx +++ b/src/views/DataView.tsx @@ -2,11 +2,15 @@ // Licensed under the MIT License. import React, { FC, useEffect, useMemo, useCallback } from 'react'; +import ReactDOM from 'react-dom'; import _ from 'lodash'; -import { Typography, Box, Link, Breadcrumbs, useTheme, Fade } from '@mui/material'; +import { Typography, Box, Link, Breadcrumbs, useTheme, Fade, IconButton, Tooltip } from '@mui/material'; import { alpha } from '@mui/material/styles'; +import { useTranslation } from 'react-i18next'; +import OpenInFullIcon from '@mui/icons-material/OpenInFull'; +import CloseFullscreenIcon from '@mui/icons-material/CloseFullscreen'; import '../scss/DataView.scss'; @@ -16,11 +20,19 @@ import { useDispatch, useSelector } from 'react-redux'; import { Type } from '../data/types'; import { SelectableDataGrid } from './SelectableDataGrid'; import { formatCellValue, getColumnAlign } from './ViewUtils'; +import { borderColor } from '../app/tokens'; export interface FreeDataViewProps { + // When true, render a maximize/restore toggle that pops the table into a + // full-canvas overlay. Used wherever the grid is shown inline (under a + // chart, or as the focused-table preview). + maximizable?: boolean; } -export const FreeDataViewFC: FC = function DataView() { +export const FreeDataViewFC: FC = function DataView({ maximizable }) { + + const { t } = useTranslation(); + const [maximized, setMaximized] = React.useState(false); const dispatch = useDispatch(); @@ -32,6 +44,7 @@ export const FreeDataViewFC: FC = function DataView() { const focusedTableId = useMemo(() => { if (!focusedId) return undefined; if (focusedId.type === 'table') return focusedId.tableId; + if (focusedId.type !== 'chart') return undefined; const chartId = focusedId.chartId; const chart = allCharts.find(c => c.id === chartId); return chart?.tableRef; @@ -108,7 +121,7 @@ export const FreeDataViewFC: FC = function DataView() { ]; }, [targetTable, rowData, conceptShelfItems]); - return ( + const grid = ( @@ -124,4 +137,77 @@ export const FreeDataViewFC: FC = function DataView() { ); + + if (!maximizable) { + return grid; + } + + const toggleButton = ( + + setMaximized(m => !m)} + sx={{ + color: 'text.secondary', + '&:hover': { color: 'primary.main', backgroundColor: 'transparent' }, + }} + > + {maximized ? : } + + + ); + + // The toggle button sits just outside the table to the right (a slim panel), + // so it never overlaps the column headers and the card keeps its original look. + // In maximized mode the surrounding overlay already provides the card frame. + const cardSx = maximized ? { overflow: 'hidden' } : { + overflow: 'hidden', + borderRadius: '8px', + border: `1px solid ${borderColor.divider}`, + transition: 'box-shadow 0.2s ease', + '&:hover': { boxShadow: '0 0 8px rgba(25, 118, 210, 0.25)' }, + }; + const framed = ( + + + {grid} + + + {toggleButton} + + + ); + + if (maximized) { + const canvas = typeof document !== 'undefined' ? document.getElementById('vis-view-canvas') : null; + const overlay = ( + <> + {/* Transparent click-catcher — click outside to restore. Scoped to the visualization view. */} + setMaximized(false)} + sx={{ position: 'absolute', inset: 0, zIndex: 1299 }} + /> + {/* Table overlay filling the visualization view. */} + + {framed} + + + ); + return ( + <> + {/* Keep the inline slot occupied so surrounding layout doesn't jump. */} + + {canvas ? ReactDOM.createPortal(overlay, canvas) : overlay} + + ); + } + + return framed; } \ No newline at end of file diff --git a/src/views/VisualizationView.tsx b/src/views/VisualizationView.tsx index 96d91d8c..7b6d18b4 100644 --- a/src/views/VisualizationView.tsx +++ b/src/views/VisualizationView.tsx @@ -932,11 +932,12 @@ export const ChartEditorFC: FC<{}> = function ChartEditorFC({}) { return sum + Math.max(80, Math.min(280, contentLen * 10)) + 60; }, ROW_ID_COL_WIDTH); const SCROLLBAR_WIDTH = 17; - const adaptiveWidth = Math.max(MIN_TABLE_WIDTH, Math.min(MAX_TABLE_WIDTH, totalColWidth + SCROLLBAR_WIDTH + 16)); + // +34px gutter so the maximize button can sit just outside the table on the right. + const adaptiveWidth = Math.max(MIN_TABLE_WIDTH, Math.min(MAX_TABLE_WIDTH, totalColWidth + SCROLLBAR_WIDTH + 16)) + 34; return ( - - + + ); })()} @@ -1096,7 +1097,7 @@ export const ChartEditorFC: FC<{}> = function ChartEditorFC({}) { , [localScaleFactor, t]); - return + return {synthesisRunning ? = function VisualizationView } return ( - + @@ -1281,18 +1282,15 @@ export const VisualizationViewFC: FC = function VisualizationView return sum + Math.max(80, Math.min(280, contentLen * 10)) + 60; }, ROW_ID_COL_WIDTH); const SCROLLBAR_WIDTH = 17; - const adaptiveWidth = Math.max(MIN_TABLE_WIDTH, Math.min(MAX_TABLE_WIDTH, totalColWidth + SCROLLBAR_WIDTH + 16)); + // +34px gutter so the maximize button can sit just outside the table on the right. + const adaptiveWidth = Math.max(MIN_TABLE_WIDTH, Math.min(MAX_TABLE_WIDTH, totalColWidth + SCROLLBAR_WIDTH + 16)) + 34; return ( - + ); })()} From a59ec9e04fd3c0c447f3fbffbe2a0f0573b7dd04 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Fri, 29 May 2026 09:25:24 -0700 Subject: [PATCH 05/29] some cleanup --- src/views/DataFormulator.tsx | 22 ++++---- src/views/DataThread.tsx | 14 ++--- src/views/SimpleChartRecBox.tsx | 98 ++++++++++++++++++++++++++------- src/views/threadLayout.ts | 39 +++++++++++++ 4 files changed, 133 insertions(+), 40 deletions(-) create mode 100644 src/views/threadLayout.ts diff --git a/src/views/DataFormulator.tsx b/src/views/DataFormulator.tsx index b340a477..90547525 100644 --- a/src/views/DataFormulator.tsx +++ b/src/views/DataFormulator.tsx @@ -40,6 +40,7 @@ import { DndProvider } from 'react-dnd' import { HTML5Backend } from 'react-dnd-html5-backend' import { toolName } from '../app/App'; import { DataThread } from './DataThread'; +import { threadPaneWidth } from './threadLayout'; import dfLogo from '../assets/df-logo.png'; import exampleImageTable from "../assets/example-image-table.png"; @@ -443,12 +444,9 @@ export const DataFormulatorFC = ({ }) => { //boxShadow: '0 0 5px rgba(0,0,0,0.1)', } - // Discrete column snapping for DataThread - const CARD_WIDTH = 220; - const CARD_GAP = 12; - const COLUMN_WIDTH = CARD_WIDTH + CARD_GAP; - const PANE_PADDING = 48; - const columnSize = (n: number) => n * COLUMN_WIDTH + PANE_PADDING; + // Discrete column snapping for DataThread. + // Column geometry is defined once in ./threadLayout and shared with + // DataThread so the pane snap points line up with the rendered columns. const allotmentRef = useRef(null); const containerRef = useRef(null); @@ -459,13 +457,13 @@ export const DataFormulatorFC = ({ }) => { let bestCols = 1; let bestDist = Infinity; for (let n = 1; n <= 3; n++) { - const dist = Math.abs(raw - columnSize(n)); + const dist = Math.abs(raw - threadPaneWidth(n)); if (dist < bestDist) { bestDist = dist; bestCols = n; } } - const snapped = columnSize(bestCols); + const snapped = threadPaneWidth(bestCols); if (Math.abs(raw - snapped) > 2) { const totalWidth = sizes.reduce((a, b) => a + b, 0); allotmentRef.current.resize([snapped, totalWidth - snapped]); @@ -545,10 +543,10 @@ export const DataFormulatorFC = ({ }) => { let newSize: number | null = null; if (prev <= 1 && threadCount > 1) { // Case 1: was 1 thread, now 2+ → expand to 2 columns - newSize = columnSize(2); + newSize = threadPaneWidth(2); } else if (prev > 1 && threadCount <= 1) { // Case 2: was 2+ threads, now 1 → shrink to 1 column - newSize = columnSize(1); + newSize = threadPaneWidth(1); } // Case 3: was 2+ threads and still 2+ → don't change (respect user's manual setting) @@ -581,7 +579,9 @@ export const DataFormulatorFC = ({ }) => { position: 'relative'}}> {tables.length > 0 ? ( - + = function ({ sx }) { // only one column fits, splitting a long thread into segments adds visual // overhead (continuation headers + ghost parents) without any layout // benefit, since the segments would just stack in the same single column. - const CARD_GAP = 12; // padding + spacing between cards in a column - const PANEL_PADDING = 16; - // 220 visual card width + 14px right gutter (CARD_CONTENT_PR) so cards - // keep their original size while gaining a right margin that balances - // the left timeline gutter. - const CARD_WIDTH = 234; - const COLUMN_WIDTH = CARD_WIDTH + CARD_GAP; - // n columns need: n*CARD_WIDTH + (n-1)*CARD_GAP + PANEL_PADDING - // Solving for n: n <= (containerWidth - PANEL_PADDING + CARD_GAP) / COLUMN_WIDTH - const fittableColumns = Math.max(1, Math.min(3, Math.floor((containerWidth - PANEL_PADDING + CARD_GAP) / COLUMN_WIDTH))); + // Column geometry (CARD_WIDTH / CARD_GAP / PANEL_PADDING) is defined once + // in ./threadLayout and shared with DataFormulator's pane snapping. + const fittableColumns = fittableThreadColumns(containerWidth); // Adaptively split long derivation chains so the resulting segments fill // the available columns evenly. See `computeSplitExtraLeaves` for the diff --git a/src/views/SimpleChartRecBox.tsx b/src/views/SimpleChartRecBox.tsx index 5e7bd63b..23b30c7c 100644 --- a/src/views/SimpleChartRecBox.tsx +++ b/src/views/SimpleChartRecBox.tsx @@ -40,7 +40,7 @@ import StopIcon from '@mui/icons-material/Stop'; import AutoGraphIcon from '@mui/icons-material/AutoGraph'; import DescriptionOutlinedIcon from '@mui/icons-material/DescriptionOutlined'; -import { UnifiedDataUploadDialog } from './UnifiedDataUploadDialog'; +import InsertDriveFileOutlinedIcon from '@mui/icons-material/InsertDriveFileOutlined'; import { borderColor, transition } from '../app/tokens'; import { Theme } from '@mui/material/styles'; import { useTranslation } from 'react-i18next'; @@ -151,7 +151,8 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ const [mentionHighlightIdx, setMentionHighlightIdx] = useState(0); const [selectedAgent, setSelectedAgent] = useState<'explore' | 'report'>('explore'); const [attachedImages, setAttachedImages] = useState([]); - const [uploadDialogOpen, setUploadDialogOpen] = useState(false); + const [attachedFiles, setAttachedFiles] = useState<{ name: string; content: string }[]>([]); + const fileInputRef = useRef(null); const agentAbortRef = useRef(null); const userChartFocusLockedRef = useRef(false); const lastAutoFocusedChartIdRef = useRef(null); @@ -296,6 +297,31 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ } }, []); + // Attach files as conversation context. Images become reference images + // (sent to the model as attachments); text-like files are read as text + // and folded into the agent prompt as context. + const handleAttachFiles = React.useCallback((fileList: FileList | null) => { + if (!fileList) return; + const MAX_TEXT_CHARS = 50000; + Array.from(fileList).forEach(file => { + if (file.type.startsWith('image/')) { + const reader = new FileReader(); + reader.onload = () => setAttachedImages(prev => [...prev, reader.result as string]); + reader.readAsDataURL(file); + } else { + const reader = new FileReader(); + reader.onload = () => { + let content = (reader.result as string) || ''; + if (content.length > MAX_TEXT_CHARS) { + content = content.slice(0, MAX_TEXT_CHARS) + '\n…[truncated]'; + } + setAttachedFiles(prev => [...prev, { name: file.name, content }]); + }; + reader.readAsText(file); + } + }); + }, []); + // Collect table IDs from root up to (and including) the focused table for agent action matching const threadTableIds = React.useMemo(() => { if (!focusedTableId) return new Set(); @@ -373,6 +399,14 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ }, displayPrompt?: string) => { if (!focusedTableId || (!clarificationContext && prompt.trim() === "")) return; + // Fold attached reference files into the prompt the agent sees, while + // keeping the timeline bubble (displayContent) clean for the user. + const fileContext = attachedFiles.length > 0 + ? '\n\n' + attachedFiles.map(f => `[Attached file: ${f.name}]\n${f.content}`).join('\n\n') + : ''; + const agentPrompt = prompt + fileContext; + const cleanDisplay = displayPrompt ?? (fileContext ? prompt : undefined); + const rootTables = tables.filter(t => t.derive === undefined || t.anchored); const currentTable = tables.find(t => t.id === focusedTableId); const priorityIds = (currentTable?.derive && !currentTable.anchored) @@ -404,8 +438,8 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ // 'clarifying' status and pendingClarification storage. if (isResume && pendingClarification?.draftId) { dispatch(dfActions.appendDraftInteraction({ draftId: pendingClarification.draftId, entry: { - from: 'user', to: 'data-agent', role: 'prompt', content: prompt, - ...(displayPrompt ? { displayContent: displayPrompt } : {}), + from: 'user', to: 'data-agent', role: 'prompt', content: agentPrompt, + ...(cleanDisplay ? { displayContent: cleanDisplay } : {}), timestamp: Date.now() }})); dispatch(dfActions.updateDraftClarification({ draftId: pendingClarification.draftId, pendingClarification: null })); @@ -552,10 +586,10 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ // backend appends it to the trajectory as a normal user message. // No special clarification payload needed. requestBody.trajectory = clarificationContext!.trajectory; - requestBody.user_question = prompt; + requestBody.user_question = agentPrompt; requestBody.completed_step_count = clarificationContext!.completedStepCount; } else { - requestBody.user_question = prompt; + requestBody.user_question = agentPrompt; if (focusedThread) requestBody.focused_thread = focusedThread; if (otherThreads) requestBody.other_threads = otherThreads; } @@ -603,13 +637,13 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ currentDraftParentTableId = existingDraft?.derive?.trigger?.tableId || null; currentDraftInteraction = [...(existingDraft?.derive?.trigger?.interaction || [])]; // The user reply was already appended above, add to local accumulator too - currentDraftInteraction.push({ from: 'user', to: 'data-agent', role: 'prompt', content: prompt, - ...(displayPrompt ? { displayContent: displayPrompt } : {}), + currentDraftInteraction.push({ from: 'user', to: 'data-agent', role: 'prompt', content: agentPrompt, + ...(cleanDisplay ? { displayContent: cleanDisplay } : {}), timestamp: Date.now() }); } else { const initialEntries: InteractionEntry[] = [ - { from: 'user', to: 'data-agent', role: 'prompt', content: prompt, - ...(displayPrompt ? { displayContent: displayPrompt } : {}), + { from: 'user', to: 'data-agent', role: 'prompt', content: agentPrompt, + ...(cleanDisplay ? { displayContent: cleanDisplay } : {}), timestamp: Date.now() } ]; createNextDraft(lastCreatedTableId || focusedTableId!, initialEntries); @@ -940,6 +974,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ clearTimeout(timeoutId); setChatPrompt(""); setAttachedImages([]); + setAttachedFiles([]); isCompleted = true; } @@ -988,6 +1023,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ clearTimeout(timeoutId); setChatPrompt(""); setAttachedImages([]); + setAttachedFiles([]); isCompleted = true; } @@ -1028,6 +1064,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ if (completionResult) { setChatPrompt(""); setAttachedImages([]); + setAttachedFiles([]); } }; @@ -1110,7 +1147,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ } } })(); - }, [focusedTableId, tables, draftNodes, activeModel, config, conceptShelfItems, dispatch, t]); + }, [focusedTableId, tables, draftNodes, activeModel, config, conceptShelfItems, dispatch, t, attachedImages, attachedFiles]); // ── Report generation via report agent ────────────────────────── @@ -1473,7 +1510,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ {/* @-mention table chips and image attachments. Skip the table-chip row entirely when there's only one root table — there's nothing else the user could @-mention, so the chip is noise. */} - {((primaryTableIds.length > 0 && rootTables.length > 1) || attachedImages.length > 0) && !isChatFormulating && ( + {((primaryTableIds.length > 0 && rootTables.length > 1) || attachedImages.length > 0 || attachedFiles.length > 0) && !isChatFormulating && ( {rootTables.length > 1 && primaryTableIds.map(id => { const tbl = tables.find(t => t.id === id); @@ -1517,6 +1554,27 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ }} /> ))} + {attachedFiles.map((file, idx) => ( + } + label={file.name} + onDelete={() => setAttachedFiles(prev => prev.filter((_, i) => i !== idx))} + sx={{ + height: 20, + fontSize: 10, + maxWidth: 160, + color: theme.palette.text.secondary, + backgroundColor: 'rgba(0,0,0,0.04)', + border: 'none', + borderRadius: '4px', + '& .MuiChip-label': { px: '4px', overflow: 'hidden', textOverflow: 'ellipsis' }, + '& .MuiChip-icon': { ml: '4px', mr: '-2px' }, + '& .MuiChip-deleteIcon': { fontSize: 12, color: theme.palette.text.disabled, mr: '2px' }, + }} + /> + ))} )} {/* @-mention dropdown */} @@ -1645,10 +1703,17 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ {/* Action buttons */} - + { handleAttachFiles(e.target.files); if (e.target) e.target.value = ''; }} + /> + { e.stopPropagation(); setUploadDialogOpen(true); }} + onClick={(e) => { e.stopPropagation(); fileInputRef.current?.click(); }} sx={{ p: 0.5, color: theme.palette.text.secondary, @@ -1762,11 +1827,6 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ {/* The input box */} {inputBox} - setUploadDialogOpen(false)} - initialTab="menu" - /> ); }; diff --git a/src/views/threadLayout.ts b/src/views/threadLayout.ts new file mode 100644 index 00000000..fa793f2c --- /dev/null +++ b/src/views/threadLayout.ts @@ -0,0 +1,39 @@ +// Single source of truth for DataThread column geometry. +// +// Both the DataThread panel (which renders the thread columns) and +// DataFormulator (which snaps the resizable Allotment pane to whole-column +// widths) must agree on these values, otherwise the pane snap points won't +// line up with the actual rendered columns. Keep all width/padding tuning +// here. + +/** Visual width of a single thread card / column (px). */ +export const CARD_WIDTH = 248; + +/** Horizontal gap between adjacent columns (px). */ +export const CARD_GAP = 8; + +/** Total horizontal padding inside the thread panel (left + right, px). */ +export const PANEL_PADDING = 32; + +/** Max number of columns the thread panel will ever lay out. */ +export const MAX_THREAD_COLUMNS = 3; + +/** + * Pixel width required to display exactly `n` columns: + * n cards + (n-1) gaps + panel padding. + */ +export const threadPaneWidth = (n: number): number => + n * CARD_WIDTH + Math.max(0, n - 1) * CARD_GAP + PANEL_PADDING; + +/** + * How many whole columns fit within `containerWidth`, clamped to + * [1, MAX_THREAD_COLUMNS]. Inverse of `threadPaneWidth`. + */ +export const fittableThreadColumns = (containerWidth: number): number => + Math.max( + 1, + Math.min( + MAX_THREAD_COLUMNS, + Math.floor((containerWidth - PANEL_PADDING + CARD_GAP) / (CARD_WIDTH + CARD_GAP)), + ), + ); From 3a23dc3039f6299845f444f42b34180ecd2cf525 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Fri, 29 May 2026 16:28:43 -0700 Subject: [PATCH 06/29] workflow design --- py-src/data_formulator/agent_config.py | 2 +- .../agents/agent_chart_insight.py | 4 +- .../agents/agent_data_loading_chat.py | 8 +- .../agents/agent_interactive_explore.py | 4 +- ...e_distill.py => agent_workflow_distill.py} | 208 ++++++++++----- py-src/data_formulator/agents/data_agent.py | 33 ++- py-src/data_formulator/app.py | 2 +- py-src/data_formulator/knowledge/store.py | 136 ++++++---- py-src/data_formulator/routes/knowledge.py | 121 +++++---- src/api/knowledgeApi.ts | 28 +- src/app/useKnowledgeStore.ts | 14 +- src/i18n/locales/en/common.json | 46 ++-- src/i18n/locales/zh/common.json | 48 ++-- src/views/DataFrameTable.tsx | 27 +- src/views/DataSourceSidebar.tsx | 2 +- src/views/DataThread.tsx | 11 - src/views/InteractionEntryCard.tsx | 5 + src/views/KnowledgePanel.tsx | 244 +++++++----------- src/views/SessionDistill.tsx | 56 ++-- src/views/SimpleChartRecBox.tsx | 46 +++- ...xperienceContext.ts => workflowContext.ts} | 4 +- .../test_agent_knowledge_integration.py | 8 +- ...ce_distill.py => test_workflow_distill.py} | 162 +++++++----- .../backend/knowledge/test_knowledge_store.py | 69 +++-- tests/backend/routes/test_knowledge_routes.py | 137 +++++----- 25 files changed, 794 insertions(+), 631 deletions(-) rename py-src/data_formulator/agents/{agent_experience_distill.py => agent_workflow_distill.py} (58%) rename src/views/{experienceContext.ts => workflowContext.ts} (98%) rename tests/backend/agents/{test_experience_distill.py => test_workflow_distill.py} (74%) diff --git a/py-src/data_formulator/agent_config.py b/py-src/data_formulator/agent_config.py index bec4c670..67dbbe31 100644 --- a/py-src/data_formulator/agent_config.py +++ b/py-src/data_formulator/agent_config.py @@ -56,7 +56,7 @@ # ── Light: single-turn extractors / classifiers / formatters ──────────── "data_load": "minimal", # one-shot type inference "data_clean": "minimal", # extract tables from text - "experience_distill": "minimal", # summarise an analysis context + "workflow_distill": "minimal", # summarise an analysis context "chart_insight": "minimal", # title + 1–3 takeaways from a chart "chart_restyle": "minimal", # apply style edits to a Vega-Lite spec "code_explanation": "minimal", # describe derived fields diff --git a/py-src/data_formulator/agents/agent_chart_insight.py b/py-src/data_formulator/agents/agent_chart_insight.py index a3ae8aba..c280efc2 100644 --- a/py-src/data_formulator/agents/agent_chart_insight.py +++ b/py-src/data_formulator/agents/agent_chart_insight.py @@ -64,7 +64,7 @@ def run(self, chart_image_base64, chart_type, field_names, input_tables=None, n= search_query = " ".join([chart_type] + field_names[:5]).strip() if search_query: relevant = self._knowledge_store.search( - search_query, categories=["experiences"], max_results=3, + search_query, categories=["workflows"], max_results=3, ) if relevant: kb_parts = ["Relevant analysis knowledge:"] @@ -72,7 +72,7 @@ def run(self, chart_image_base64, chart_type, field_names, input_tables=None, n= kb_parts.append(f"- {item['title']}: {item['snippet'][:200]}") context_parts.append("\n".join(kb_parts)) except Exception: - logger.warning("Failed to search knowledge experiences", exc_info=True) + logger.warning("Failed to search knowledge workflows", exc_info=True) context = "\n".join(context_parts) diff --git a/py-src/data_formulator/agents/agent_data_loading_chat.py b/py-src/data_formulator/agents/agent_data_loading_chat.py index 61d3a0e6..55f2640a 100644 --- a/py-src/data_formulator/agents/agent_data_loading_chat.py +++ b/py-src/data_formulator/agents/agent_data_loading_chat.py @@ -1292,7 +1292,7 @@ def _build_system_prompt(self, last_user_text: str = ""): """Build the system prompt with current workspace context. *last_user_text* is used to search the knowledge store for - experiences relevant to the user's current request. Falls back + workflows relevant to the user's current request. Falls back to a generic query when empty. """ table_names = "none" @@ -1324,7 +1324,7 @@ def _build_system_prompt(self, last_user_text: str = ""): if self._knowledge_store: prompt += self._knowledge_store.format_rules_block() - # Inject relevant experiences from knowledge store + # Inject relevant workflows from knowledge store if self._knowledge_store: try: search_query = ( @@ -1334,7 +1334,7 @@ def _build_system_prompt(self, last_user_text: str = ""): ) relevant = self._knowledge_store.search( search_query, - categories=["experiences"], + categories=["workflows"], max_results=3, ) if relevant: @@ -1343,7 +1343,7 @@ def _build_system_prompt(self, last_user_text: str = ""): knowledge_block += f"\n### {item['title']}\n{item['snippet']}\n" prompt += "\n\n" + knowledge_block except Exception: - logger.warning("Failed to search knowledge experiences", exc_info=True) + logger.warning("Failed to search knowledge workflows", exc_info=True) if self.language_instruction: prompt += "\n\n" + self.language_instruction diff --git a/py-src/data_formulator/agents/agent_interactive_explore.py b/py-src/data_formulator/agents/agent_interactive_explore.py index 67847ec2..0f5f90fb 100644 --- a/py-src/data_formulator/agents/agent_interactive_explore.py +++ b/py-src/data_formulator/agents/agent_interactive_explore.py @@ -162,7 +162,7 @@ def run(self, input_tables, start_question=None, if start_question: context += f"\n\n[START QUESTION]\n\n{start_question}" - # ── Inject relevant experiences from knowledge store ────────── + # ── Inject relevant workflows from knowledge store ────────── if self._knowledge_store: try: query = start_question or "" @@ -170,7 +170,7 @@ def run(self, input_tables, start_question=None, search_query = " ".join([query] + table_names[:5]).strip() if search_query: relevant = self._knowledge_store.search( - search_query, categories=["experiences"], max_results=3, + search_query, categories=["workflows"], max_results=3, ) if relevant: knowledge_block = "[RELEVANT KNOWLEDGE]\n" diff --git a/py-src/data_formulator/agents/agent_experience_distill.py b/py-src/data_formulator/agents/agent_workflow_distill.py similarity index 58% rename from py-src/data_formulator/agents/agent_experience_distill.py rename to py-src/data_formulator/agents/agent_workflow_distill.py index cc738495..3f3d9c6d 100644 --- a/py-src/data_formulator/agents/agent_experience_distill.py +++ b/py-src/data_formulator/agents/agent_workflow_distill.py @@ -1,17 +1,17 @@ # Copyright (c) Microsoft Corporation. # Licensed under the MIT License. -"""Experience distillation agent — extracts reusable knowledge from analysis context. +"""Workflow distillation agent — extracts a replayable workflow from analysis context. Given a user-visible analysis context (timeline of events) plus an optional user instruction, this agent calls an LLM to produce a structured Markdown -experience document with YAML front matter suitable for storage in the +workflow document with YAML front matter suitable for storage in the knowledge base. Usage:: - agent = ExperienceDistillAgent(client) - md_content = agent.run(experience_context, user_instruction="...") + agent = WorkflowDistillAgent(client) + md_content = agent.run(workflow_context, user_instruction="...") """ from __future__ import annotations @@ -25,18 +25,19 @@ logger = logging.getLogger(__name__) -_AGENT_ID = "experience_distill" +_AGENT_ID = "workflow_distill" SYSTEM_PROMPT = """\ -You are a knowledge distiller. Given the chronological events of a data -analysis session plus an optional user instruction, write a short reusable -Markdown note that will help with similar future tasks. +You are a workflow distiller. Given the chronological events of a data +analysis session plus an optional user instruction, extract a short, +**replayable workflow** that captures *what the user wanted and got* — so +the same analysis can be reproduced later on a similarly-shaped dataset. The session contains one or more threads (separate analysis branches in the same session) each rendered under a `### Thread N` header. When -multiple threads are provided, synthesise lessons that hold across them -— do NOT enumerate per-thread. +multiple threads are provided, merge them into one coherent ordered +workflow — do NOT enumerate per-thread. The events use three types: - `message` — directed speech, formatted as `[→/] `. @@ -46,56 +47,136 @@ (followed by columns, row count, sample, and code). - `create_chart` — a chart emitted on a table (mark + encoding summary). -If a user instruction is provided, focus the note on that instruction. -Otherwise, distill the most transferable methodology from the events. +Your job is to recover the **ordered list of requests** the user actually +wanted, and the outputs (tables/charts) they ended up keeping. Beyond the +concrete steps, also distill the analysis at TWO levels of abstraction so +it can be reused later: +- **Adapting to similar data** (concrete) — how to rerun essentially the + same analysis on a near-identical dataset, e.g. the business report for + a different month, region, or product line. Same shape and intent, only + the specific inputs/filters change. +- **Generalizing to other data** (abstract, dataset-agnostic) — the + underlying analytical pattern, independent of this domain: the kinds of + questions, computations, and charts involved, phrased so they transfer + to a different domain or a differently-shaped dataset. + +CRITICAL extraction rules — keep only what the user wanted and got: +- Each step = one user request, written in plain language. Say BOTH the + question being explored AND what was produced to answer it — including + the chart that was created and the key fields it uses (e.g. "Ask how + sales trend over time, and plot monthly total sales as a line chart"; + "Compare regions by breaking revenue down per region as a sorted bar + chart"). Order them as the analysis progressed. +- DROP corrective back-and-forth. If the user changed their mind + ("no, it should be…", "actually use median instead"), keep ONLY the + final resolved intent — not the wrong first attempt or the correction. +- DROP abandoned work. If a chart or table was created and then deleted + or never kept, leave it out entirely. +- DROP mechanics. Do NOT include error-repair loops, dtype fixes, tool + call noise, or low-level code. Describe intent, not implementation. +- Do NOT lean on code or exact column names unless a name is essential to + the request's meaning. Keep steps dataset-agnostic where possible so + they replay on a new slice of similar data. +- Capture genuine gotchas separately as short notes (advisory warnings to + carry forward), NOT as steps to re-perform. + +If a user instruction is provided, let it steer what to keep or emphasise. Output format (Markdown with YAML front matter, nothing else): ``` --- -subtitle: -tags: [] +subtitle: +filename: created: updated: source: distill source_context: --- -## When to Use - - -## Method - - -## Pitfalls & Tips - +## Goal + + +## Steps +1. +2. +3. <…> + +## Adapting to similar data + + +## Generalizing to other data + + +## Notes + ``` Rules: -- Subtitle must be a short, scannable noun phrase (3-8 words) that captures - the technique or pattern. The hosting application prefixes it with the - session name to form the full title (e.g. "Experience from : "), - so do NOT include the session name in the subtitle. Do NOT pack scenario, - takeaway, and steps into the subtitle — leave details for `## When to Use` - and `## Method`. - Good: "Year-over-year volatility comparison". "Repairing pandas dtype mismatches". - Bad: "Time series analysis workflow: aggregate, visualize trends, quantify YoY spikes, and compare volatility across periods". -- Focus on *transferable* methods and caveats, not case-specific details. -- Keep the body under 500 words. -- No raw data, PII, secrets, or specific values unless they show a universal pattern. -- Write the subtitle, headings, body, and tags in {output_language}. +- Subtitle must DESCRIBE what the workflow is about in PLAIN LANGUAGE that + a non-expert can fully understand at a glance, so they can decide + whether to replay it on new data. Favor clarity over brevity: it can be + a full sentence (up to ~25 words) if that makes the analysis genuinely + understandable. Write it like you would explain the analysis to a + colleague in one breath, covering the subject and the main thing you do + with it. The hosting application uses this subtitle directly as the + workflow's display title, so make it self-contained and do NOT prefix it + with the session name. + - Start with a concrete action verb (Plot, Compare, Break down, Rank, + Track, Summarize, Find…). + - Name the real-world subject in everyday words (sales, revenue, + customers, events), NOT the internal mechanics or derived-column + names you happened to create. + - AVOID abstract or technical jargon and invented noun-phrases + ("deltas", "composition", "window", "distribution shift"). If a + technique matters, phrase it plainly ("change from one period to the + next" instead of "deltas"). + Good: "Plot monthly sales over time and compare each year against the + previous one to spot volatile periods". + "Break revenue down by region and show how each region + contributes to the total as a stacked area chart". + "Track how many events happen in each time window and what kinds + of events make up each window". + Bad: "Time series analysis". "Data workflow". "Chart exploration". + "Event window deltas with composition". "Distribution shift inspection". +- Filename must be a SHORT (2-5 word) lowercase name for the file — just + the core subject and action, e.g. "monthly sales trend", "region revenue + breakdown". No dates, no file extension, no session name. It is only + used to name the file on disk; the descriptive subtitle is what users see. +- Steps must be ordered and reproducible. Each step should make clear the + question being explored and the chart/output produced to answer it. +- "Adapting to similar data" stays close to this analysis (same domain, + same shape) — only the concrete inputs change. "Generalizing to other + data" must be domain-neutral: strip out this dataset's subject matter and + describe only the transferable analytical pattern (question types, + computations, chart kinds). Do NOT just repeat the steps in either + section; add genuine reuse guidance. Keep each section brief. +- Be as long as the analysis needs — do not omit meaningful steps, + questions, or charts just to stay short. Stay focused, but completeness + matters more than brevity. +- No raw data, PII, secrets, or specific values unless essential to a request. +- Write the subtitle, headings, and body in {output_language}. YAML front-matter keys stay in English. {language_instruction} """ -class ExperienceDistillAgent: - """Distills analysis context into a reusable experience document.""" - # Language display names for experience-specific prompts +class WorkflowDistillAgent: + """Distills analysis context into a reusable workflow document.""" + + # Language display names for workflow-specific prompts _LANG_NAMES: dict[str, str] = { "zh": "Simplified Chinese (简体中文)", "ja": "Japanese (日本語)", @@ -121,7 +202,7 @@ def __init__( self.timeout_seconds = int(timeout_seconds) if timeout_seconds else self.DEFAULT_TIMEOUT def run(self, context: dict[str, Any], user_instruction: str = "") -> str: - """Distill an experience document from user-visible session context.""" + """Distill a workflow document from user-visible session context.""" summary = self._extract_context_summary(context) today = datetime.now(timezone.utc).strftime("%Y-%m-%d") context_id = str(context.get("context_id", "") or "") @@ -130,7 +211,7 @@ def run(self, context: dict[str, Any], user_instruction: str = "") -> str: instruction_block = ( f"\n[USER INSTRUCTION]\n{user_instruction.strip()}\n" - f"Focus the distilled experience on the above instruction.\n" + f"Focus the distilled workflow on the above instruction.\n" ) if user_instruction and user_instruction.strip() else "" workspace_block = ( @@ -158,9 +239,11 @@ def run(self, context: dict[str, Any], user_instruction: str = "") -> str: {"role": "user", "content": user_msg}, ] - from data_formulator.knowledge.store import KNOWLEDGE_LIMITS + from data_formulator.knowledge.store import KNOWLEDGE_LIMITS, WORKFLOW_HARD_MAX content = self._call_with_length_retry( - messages, KNOWLEDGE_LIMITS.get("experiences", 2000), + messages, + KNOWLEDGE_LIMITS.get("workflows", 6000), + WORKFLOW_HARD_MAX, ) if not content.strip().startswith("---"): @@ -182,7 +265,7 @@ def _prompt_format_kwargs(self) -> dict[str, str]: lang_block = ( f"[LANGUAGE INSTRUCTION]\n" f"The user's language is **{display_name}**.\n" - f"Write the title, all section headings, all body text, and tags " + f"Write the title, all section headings, and all body text " f"in {display_name}. YAML front-matter keys stay in English." ) return { @@ -199,39 +282,43 @@ def _prompt_format_kwargs(self) -> dict[str, str]: def _call_with_length_retry( self, messages: list[dict], - body_limit: int, + soft_limit: int, + hard_limit: int, ) -> str: - """Call LLM and retry once if the body exceeds *body_limit* characters. + """Call the LLM, nudging it to stay near *soft_limit* characters. - If the retry *still* overshoots, hard-truncate the body so the - document is saved instead of the entire distillation being lost. + ``soft_limit`` is advisory guidance: if the first response overshoots + it we retry once asking the model to condense. We only ever + hard-truncate at ``hard_limit`` — a much larger safety ceiling — so + rich, multi-section workflows are kept intact while runaway output + is still bounded. """ from data_formulator.knowledge.store import parse_front_matter content = self._call_llm(messages) _, body = parse_front_matter(content) - if len(body.strip()) <= body_limit: + if len(body.strip()) <= soft_limit: return content - retry_target = max(body_limit - self.RETRY_MARGIN, 1) + retry_target = max(soft_limit - self.RETRY_MARGIN, 1) logger.info( - "Distilled content too long (%d > %d), retrying with condensation prompt (target ≤ %d)", - len(body.strip()), body_limit, retry_target, + "Distilled content over soft target (%d > %d), retrying with condensation prompt (target ≤ %d)", + len(body.strip()), soft_limit, retry_target, ) messages = messages + [ {"role": "assistant", "content": content}, {"role": "user", "content": ( - f"Your output body is {len(body.strip())} characters, which exceeds " - f"the limit of {body_limit}. Please condense the document to fit " - f"within {retry_target} characters while keeping the most important " - f"insights. Output ONLY the revised Markdown document." + f"Your output body is {len(body.strip())} characters, which is " + f"longer than ideal. Please tighten the document to around " + f"{retry_target} characters while keeping the most important " + f"insights and all sections. Output ONLY the revised Markdown document." )}, ] retried = self._call_llm(messages) - # Hard-trim if the retry still overshoots — better a slightly - # truncated experience than a save failure. - return self._truncate_body_to_limit(retried, body_limit) + # Hard-trim only if the retry blows past the absolute ceiling — + # better a slightly truncated workflow than a save failure. + return self._truncate_body_to_limit(retried, hard_limit) @classmethod def _truncate_body_to_limit(cls, content: str, body_limit: int) -> str: @@ -385,7 +472,7 @@ def _render_events(cls, events: list[Any]) -> str: return "\n".join(parts) if parts else "(empty context)" def _call_llm(self, messages: list[dict]) -> str: - """Single LLM call to generate the experience document.""" + """Single LLM call to generate the workflow document.""" resp = self.client.get_completion( messages, reasoning_effort=reasoning_effort_for(_AGENT_ID, self.client.model), timeout=self.timeout_seconds, ) @@ -401,7 +488,6 @@ def _add_fallback_front_matter( header = ( f"---\ntitle: {title}\n" - f"tags: []\n" f"created: {today}\n" f"updated: {today}\n" f"source: distill\n" diff --git a/py-src/data_formulator/agents/data_agent.py b/py-src/data_formulator/agents/data_agent.py index 8e9cd39a..9a9d10b2 100644 --- a/py-src/data_formulator/agents/data_agent.py +++ b/py-src/data_formulator/agents/data_agent.py @@ -153,7 +153,7 @@ def _rescue_validate_action(data: dict) -> list[str]: "function": { "name": "search_knowledge", "description": ( - "Search the user's knowledge base (rules, experiences) " + "Search the user's knowledge base (rules, workflows) " "for relevant entries. Returns title, category, snippet, and " "path for each match. Use read_knowledge to get full content." ), @@ -168,7 +168,7 @@ def _rescue_validate_action(data: dict) -> list[str]: "type": "array", "items": { "type": "string", - "enum": ["rules", "experiences"], + "enum": ["rules", "workflows"], }, "description": "Optional: limit search to specific categories.", }, @@ -190,7 +190,7 @@ def _rescue_validate_action(data: dict) -> list[str]: "properties": { "category": { "type": "string", - "enum": ["rules", "experiences"], + "enum": ["rules", "workflows"], "description": "Knowledge category.", }, "path": { @@ -224,7 +224,7 @@ def _rescue_validate_action(data: dict) -> list[str]: - **inspect_source_data(table_names)** — get schema, stats, and sample rows for source tables (cheaper than explore for basic inspection). - **search_knowledge(query, categories?)** — search the user's knowledge base - (rules, experiences) for relevant entries. + (rules, workflows) for relevant entries. - **read_knowledge(category, path)** — read the full content of a knowledge entry. You analyse data that is **already in the workspace**. If the user's @@ -1379,14 +1379,14 @@ def _build_initial_messages( if peripheral_block: user_content += f"{peripheral_block}\n\n" - # Search and inject relevant knowledge (experiences + non-alwaysApply rules) + # Search and inject relevant knowledge (workflows + non-alwaysApply rules) table_names = [t.get("name", "") for t in input_tables if t.get("name")] relevant_knowledge = self._search_relevant_knowledge(user_question, table_names) - # Always include the experience distilled from the active workspace + # Always include the workflow distilled from the active workspace # (design-docs/24 §3.6) so the session has stable working memory # across turns regardless of search relevance. - session_exp = self._load_active_session_experience() + session_exp = self._load_active_session_workflow() if session_exp: existing_paths = { (item["category"], item["path"]) for item in relevant_knowledge @@ -1891,7 +1891,7 @@ def _search_relevant_knowledge( table_names: list[str], max_items: int = 5, ) -> list[dict[str, Any]]: - """Search experiences and non-alwaysApply rules relevant to the current session. + """Search workflows and non-alwaysApply rules relevant to the current session. Uses the user question as the search query and passes table names separately for tag-overlap boosting. alwaysApply rules are @@ -1904,7 +1904,7 @@ def _search_relevant_knowledge( try: results = self._knowledge_store.search( user_question, - categories=["rules", "experiences"], + categories=["rules", "workflows"], max_results=max_items, table_names=table_names[:5], ) @@ -1913,11 +1913,11 @@ def _search_relevant_knowledge( logger.warning("Failed to search knowledge", exc_info=True) return [] - def _load_active_session_experience(self) -> dict[str, Any] | None: - """Return the experience distilled from the active workspace, if any. + def _load_active_session_workflow(self) -> dict[str, Any] | None: + """Return the workflow distilled from the active workspace, if any. The session-scoped distillation flow (design-docs/24) writes one - experience per workspace, stamped with ``source_workspace_id``. + workflow per workspace, stamped with ``source_workspace_id``. We always inject that file into the agent's context so the agent has stable working memory for the active session in addition to whatever the relevance search picked. @@ -1932,14 +1932,14 @@ def _load_active_session_experience(self) -> dict[str, Any] | None: if not ws_id: return None try: - entry = self._knowledge_store.find_experience_by_workspace_id(ws_id) + entry = self._knowledge_store.find_workflow_by_workspace_id(ws_id) except Exception: - logger.warning("find_experience_by_workspace_id failed", exc_info=True) + logger.warning("find_workflow_by_workspace_id failed", exc_info=True) return None if not entry: return None try: - content = self._knowledge_store.read("experiences", entry["path"]) + content = self._knowledge_store.read("workflows", entry["path"]) except Exception: return None from data_formulator.knowledge.store import parse_front_matter @@ -1948,9 +1948,8 @@ def _load_active_session_experience(self) -> dict[str, Any] | None: if not snippet: return None return { - "category": "experiences", + "category": "workflows", "title": entry.get("title", entry.get("path", "")), - "tags": entry.get("tags", []), "path": entry["path"], "snippet": snippet, "source": entry.get("source", "distill"), diff --git a/py-src/data_formulator/app.py b/py-src/data_formulator/app.py index 47d2bda8..ef9dd4cb 100644 --- a/py-src/data_formulator/app.py +++ b/py-src/data_formulator/app.py @@ -219,7 +219,7 @@ def _register_blueprints(): from data_formulator.routes.credentials import credential_bp app.register_blueprint(credential_bp) - # Register knowledge management API (rules, skills, experiences) + # Register knowledge management API (rules, skills, workflows) from data_formulator.routes.knowledge import knowledge_bp app.register_blueprint(knowledge_bp) diff --git a/py-src/data_formulator/knowledge/store.py b/py-src/data_formulator/knowledge/store.py index 0b290093..08463437 100644 --- a/py-src/data_formulator/knowledge/store.py +++ b/py-src/data_formulator/knowledge/store.py @@ -1,10 +1,10 @@ # Copyright (c) Microsoft Corporation. # Licensed under the MIT License. -"""Knowledge store — manages user knowledge files (rules, experiences). +"""Knowledge store — manages user knowledge files (rules, workflows). Each user has a ``knowledge/`` directory under their home with two -sub-directories: ``rules`` and ``experiences``. Every knowledge entry is a +sub-directories: ``rules`` and ``workflows``. Every knowledge entry is a Markdown file with YAML front matter. All file I/O is routed through :class:`ConfinedDir` for path safety. @@ -12,7 +12,7 @@ Directory depth constraints: - ``rules``: flat — only files directly under ``rules/`` (1 path part) -- ``experiences``: one level of sub-directories (up to 2 path parts) +- ``workflows``: one level of sub-directories (up to 2 path parts) """ from __future__ import annotations @@ -27,19 +27,27 @@ logger = logging.getLogger(__name__) -VALID_CATEGORIES = frozenset({"rules", "experiences"}) +VALID_CATEGORIES = frozenset({"rules", "workflows"}) _MAX_DEPTH = { "rules": 1, - "experiences": 2, # one sub-dir: "category/file.md" + "workflows": 2, # one sub-dir: "category/file.md" } KNOWLEDGE_LIMITS: dict[str, int] = { "rule_description_max": 100, "rules": 350, - "experiences": 2000, + # Soft length guidance for distilled workflows: the target the distill + # agent aims for, NOT a hard cap. Workflows may exceed it when an + # analysis genuinely needs the room (e.g. multiple abstraction levels). + # Writes are only rejected past WORKFLOW_HARD_MAX below. + "workflows": 6000, } +# Absolute safety ceiling for a workflow body. Guards against runaway LLM +# output while still letting rich, multi-section workflows through. +WORKFLOW_HARD_MAX: int = 24000 + # --------------------------------------------------------------------------- # Tokenization helpers for improved search scoring # --------------------------------------------------------------------------- @@ -151,14 +159,13 @@ class KnowledgeItemMeta: """ __slots__ = ( - "title", "tags", "source", "created", "description", "always_apply", + "title", "source", "created", "description", "always_apply", "source_workspace_id", "source_workspace_name", ) def __init__( self, title: str, - tags: list[str], source: str, created: str, description: str, @@ -167,7 +174,6 @@ def __init__( source_workspace_name: str = "", ): self.title = title - self.tags = tags self.source = source self.created = created self.description = description @@ -181,14 +187,6 @@ def from_raw(cls, meta: dict[str, Any], fallback_stem: str = "") -> "KnowledgeIt title = meta.get("title", fallback_stem) title = str(title) if title is not None else fallback_stem - raw_tags = meta.get("tags", []) - if isinstance(raw_tags, list): - tags = [str(t) for t in raw_tags] - elif raw_tags is None: - tags = [] - else: - tags = [str(raw_tags)] - source = str(meta.get("source", "manual") or "manual") created = str(meta.get("created", "") or "") description = str(meta.get("description", "") or "") @@ -198,7 +196,6 @@ def from_raw(cls, meta: dict[str, Any], fallback_stem: str = "") -> "KnowledgeIt return cls( title=title, - tags=tags, source=source, created=created, description=description, @@ -246,26 +243,64 @@ class KnowledgeStore: store = KnowledgeStore(user_home) items = store.list_all("rules") - content = store.read("experiences", "data-cleaning/handle-missing.md") + content = store.read("workflows", "data-cleaning/handle-missing.md") store.write("rules", "date-format.md", md_content) store.delete("rules", "date-format.md") - results = store.search("ROI", categories=["rules", "experiences"]) + results = store.search("ROI", categories=["rules", "workflows"]) """ def __init__(self, user_home: Path | str) -> None: user_home = Path(user_home) self._root = ConfinedDir(user_home / "knowledge", mkdir=True) + self._migrate_experiences_to_workflows() self._jails: dict[str, ConfinedDir] = { "rules": ConfinedDir(self._root.root / "rules", mkdir=True), - "experiences": ConfinedDir(self._root.root / "experiences", mkdir=True), + "workflows": ConfinedDir(self._root.root / "workflows", mkdir=True), } self._migrate_flat() # -- migration --------------------------------------------------------- + def _migrate_experiences_to_workflows(self) -> None: + """Move legacy ``experiences/`` files into ``workflows/`` (one-time). + + The feature was renamed from "experiences" to "workflows"; existing + users have files under ``knowledge/experiences/``. Move them so the + rename is transparent. + """ + old_root = self._root.root / "experiences" + if not old_root.is_dir(): + return + new_root = self._root.root / "workflows" + new_root.mkdir(parents=True, exist_ok=True) + for md_file in list(old_root.rglob("*.md")): + rel = md_file.relative_to(old_root) + dest = new_root / rel + dest.parent.mkdir(parents=True, exist_ok=True) + if dest.exists(): + stem = rel.stem + suffix_n = 1 + while dest.exists(): + dest = dest.parent / f"{stem}-{suffix_n}.md" + suffix_n += 1 + try: + md_file.rename(dest) + logger.info("Migrated experiences/%s → workflows/%s", rel, dest.name) + except Exception: + logger.warning("Failed to migrate experience file %s", md_file, exc_info=True) + # Remove the now-empty legacy tree (best effort) + try: + for sub in sorted(old_root.rglob("*"), reverse=True): + if sub.is_dir() and not any(sub.iterdir()): + sub.rmdir() + if not any(old_root.iterdir()): + old_root.rmdir() + except Exception: + logger.warning("Failed to clean up legacy experiences dir", exc_info=True) + def _migrate_flat(self) -> None: - """Move any experiences/subdir/file.md → experiences/file.md (one-time migration).""" - exp_root = self._jails["experiences"].root + """Move any workflows/subdir/file.md → workflows/file.md (one-time migration).""" + exp_root = self._jails["workflows"].root for md_file in list(exp_root.rglob("*.md")): rel = md_file.relative_to(exp_root) if len(rel.parts) <= 1: @@ -285,9 +320,9 @@ def _migrate_flat(self) -> None: parent = md_file.parent if parent != exp_root and not any(parent.iterdir()): parent.rmdir() - logger.info("Migrated knowledge experience %s → %s", rel, dest.name) + logger.info("Migrated knowledge workflow %s → %s", rel, dest.name) except Exception: - logger.warning("Failed to migrate experience file %s", md_file, exc_info=True) + logger.warning("Failed to migrate workflow file %s", md_file, exc_info=True) # -- path validation --------------------------------------------------- @@ -326,7 +361,7 @@ def _jail(self, category: str) -> ConfinedDir: def list_all(self, category: str) -> list[dict[str, Any]]: """List all knowledge entries in *category*. - Returns a list of dicts with ``title``, ``tags``, ``path``, + Returns a list of dicts with ``title``, ``path``, ``source``, and ``created`` parsed from front matter. For rules, also includes ``description`` and ``alwaysApply``. """ @@ -345,7 +380,6 @@ def list_all(self, category: str) -> list[dict[str, Any]]: rel = str(md_file.relative_to(jail.root)).replace("\\", "/") item: dict[str, Any] = { "title": km.title, - "tags": km.tags, "path": rel, "source": km.source, "created": km.created, @@ -353,9 +387,9 @@ def list_all(self, category: str) -> list[dict[str, Any]]: if category == "rules": item["description"] = km.description item["alwaysApply"] = km.always_apply - if category == "experiences": + if category == "workflows": # Surface session-distillation provenance so the frontend can - # find an existing session experience by workspace id + # find an existing session workflow by workspace id # without re-reading every file. See design-docs/24. if km.source_workspace_id: item["sourceWorkspaceId"] = km.source_workspace_id @@ -394,7 +428,15 @@ def write(self, category: str, path: str, content: str) -> Path: body_limit = KNOWLEDGE_LIMITS.get(category) if body_limit is not None: body_len = len(body.strip()) - if body_len > body_limit: + if category == "workflows": + # Soft guidance: the body_limit is a target the distill agent + # aims for, not a hard cap. Only reject far past the ceiling. + if body_len > WORKFLOW_HARD_MAX: + raise ValueError( + f"workflows body exceeds {WORKFLOW_HARD_MAX} characters " + f"(got {body_len})" + ) + elif body_len > body_limit: raise ValueError( f"{category} body exceeds {body_limit} characters " f"(got {body_len})" @@ -407,12 +449,12 @@ def delete(self, category: str, path: str) -> None: self.validate_path(category, path) self._jail(category).unlink(path) - # -- session experience helpers ---------------------------------------- + # -- session workflow helpers ---------------------------------------- - def find_experience_by_workspace_id( + def find_workflow_by_workspace_id( self, workspace_id: str, ) -> dict[str, Any] | None: - """Return the experience entry whose front matter records this workspace id. + """Return the workflow entry whose front matter records this workspace id. Used by the session-scoped distillation flow (design-docs/24) to upsert: when re-distilling the same session, overwrite the same @@ -421,11 +463,11 @@ def find_experience_by_workspace_id( if not workspace_id or not workspace_id.strip(): return None try: - for item in self.list_all("experiences"): + for item in self.list_all("workflows"): if item.get("sourceWorkspaceId") == workspace_id: return item except Exception: - logger.warning("find_experience_by_workspace_id failed", exc_info=True) + logger.warning("find_workflow_by_workspace_id failed", exc_info=True) return None # -- alwaysApply rules helper ------------------------------------------ @@ -511,12 +553,13 @@ def search( """Search across knowledge categories. Tokenizes *query* into keywords and scores each entry using - multi-field weighted matching (title > tags > filename > body). - Whole-string exact matches and table-name / tag overlaps receive + multi-field weighted matching (title > filename > body). + Whole-string exact matches and table-name overlaps receive additional bonuses. Non-manual sources are slightly discounted. *table_names* (optional) are table names from the current session; - when a table name appears in an entry's tags the entry is boosted. + when a table name appears in an entry's title or body the entry is + boosted. """ if not query or not query.strip(): return [] @@ -542,7 +585,7 @@ def search( continue score = self._match_score( - q, km.title, km.tags, md_file.stem, body[:200], + q, km.title, md_file.stem, body[:200], source=km.source, table_names=table_names, ) if score <= 0: @@ -552,7 +595,6 @@ def search( scored.append((score, { "category": cat, "title": km.title, - "tags": km.tags, "path": rel, "snippet": body[:500].strip(), "source": km.source, @@ -565,7 +607,6 @@ def search( def _match_score( query: str, title: str, - tags: list[str], stem: str, body_prefix: str, *, @@ -589,13 +630,10 @@ def _match_score( title_l = title.lower() stem_l = stem.lower() body_l = body_prefix.lower() - tags_l = [t.lower() for t in tags] for token in tokens: if token in title_l: score += 100 / n - if any(token in tl for tl in tags_l): - score += 50 / n if token in stem_l: score += 30 / n if token in body_l: @@ -604,14 +642,14 @@ def _match_score( # Whole-string bonus (handles short queries like "ROI") if q and q in title.lower(): score += 50 - if q and any(q in t.lower() for t in tags): - score += 50 - # Table-name → tag overlap bonus + # Table-name overlap bonus (title / body) if table_names: - tags_l_set = {t.lower() for t in tags} + title_l = title.lower() + body_l = body_prefix.lower() for tn in table_names: - if any(tn.lower() in tl for tl in tags_l_set): + tnl = tn.lower() + if tnl in title_l or tnl in body_l: score += 30 # Non-manual source slight discount diff --git a/py-src/data_formulator/routes/knowledge.py b/py-src/data_formulator/routes/knowledge.py index 901024c1..1a458ba9 100644 --- a/py-src/data_formulator/routes/knowledge.py +++ b/py-src/data_formulator/routes/knowledge.py @@ -1,7 +1,7 @@ # Copyright (c) Microsoft Corporation. # Licensed under the MIT License. -"""Knowledge management API — CRUD + search + experience distillation. +"""Knowledge management API — CRUD + search + workflow distillation. All endpoints use ``POST`` with JSON body. Access is scoped to the current user via ``get_identity_id()`` and confined via ``ConfinedDir``. @@ -155,44 +155,44 @@ def knowledge_search(): return json_ok({"results": results}) -# ── distill experience ──────────────────────────────────────────────────── +# ── distill workflow ──────────────────────────────────────────────────── -@knowledge_bp.route("/distill-experience", methods=["POST"]) -def distill_experience(): - """Distill user-visible analysis context into a reusable experience. +@knowledge_bp.route("/distill-workflow", methods=["POST"]) +def distill_workflow(): + """Distill user-visible analysis context into a reusable workflow. Session-scoped payload (design-docs/24): - ``experience_context`` carries a list of ``threads`` (one per leaf + ``workflow_context`` carries a list of ``threads`` (one per leaf derived table the user has on screen), each with its own chronological ``events`` array. ``workspace_id`` + ``workspace_name`` bind the resulting file to the active session so re-distilling upserts the same file. - Required body fields: ``experience_context`` and ``model``. + Required body fields: ``workflow_context`` and ``model``. Optional: ``user_instruction`` (natural-language focus hint for the LLM), - ``category_hint`` (sub-directory under experiences/). + ``category_hint`` (sub-directory under workflows/). """ data = request.get_json(silent=True) or {} - experience_context = data.get("experience_context") - if not isinstance(experience_context, dict): - raise AppError(ErrorCode.INVALID_REQUEST, "'experience_context' is required") + workflow_context = data.get("workflow_context") + if not isinstance(workflow_context, dict): + raise AppError(ErrorCode.INVALID_REQUEST, "'workflow_context' is required") - threads = experience_context.get("threads") + threads = workflow_context.get("threads") if not isinstance(threads, list) or not threads: raise AppError( ErrorCode.INVALID_REQUEST, - "'experience_context.threads' is required and must be a non-empty list", + "'workflow_context.threads' is required and must be a non-empty list", ) - workspace_id_raw = experience_context.get("workspace_id", "") + workspace_id_raw = workflow_context.get("workspace_id", "") workspace_id = workspace_id_raw.strip() if isinstance(workspace_id_raw, str) else "" - workspace_name_raw = experience_context.get("workspace_name", "") + workspace_name_raw = workflow_context.get("workspace_name", "") workspace_name = workspace_name_raw.strip() if isinstance(workspace_name_raw, str) else "" if not workspace_id or not workspace_name: raise AppError( ErrorCode.INVALID_REQUEST, - "'experience_context.workspace_id' and 'workspace_name' are required", + "'workflow_context.workspace_id' and 'workspace_name' are required", ) model_config = data.get("model") @@ -215,53 +215,55 @@ def distill_experience(): # Build client and run distillation from data_formulator.routes.agents import get_client, _get_ui_lang - from data_formulator.agents.agent_experience_distill import ExperienceDistillAgent + from data_formulator.agents.agent_workflow_distill import WorkflowDistillAgent client = get_client(model_config) - agent = ExperienceDistillAgent( + agent = WorkflowDistillAgent( client=client, language_code=_get_ui_lang(), timeout_seconds=timeout_seconds, ) try: - md_content = agent.run(experience_context, user_instruction=user_instruction) + md_content = agent.run(workflow_context, user_instruction=user_instruction) except Exception as exc: - logger.warning("Experience distillation LLM call failed: %s", type(exc).__name__) + logger.warning("Workflow distillation LLM call failed: %s", type(exc).__name__) from data_formulator.error_handler import classify_and_wrap_llm_error raise classify_and_wrap_llm_error(exc) from exc - # Save to knowledge/experiences/ + # Save to knowledge/workflows/ store = KnowledgeStore(user_home) - # Bind the file to the workspace, override title to - # "Experience from : ", and upsert below. - md_content = _apply_session_front_matter(md_content, workspace_id, workspace_name) + # Bind the file to the workspace, set the title to the agent-generated + # descriptive subtitle, and upsert below. + md_content, title_core, filename_hint = _apply_session_front_matter( + md_content, workspace_id, workspace_name, + ) - filename = _experience_filename(workspace_name) + filename = _workflow_filename(filename_hint or title_core or workspace_name) rel_path = f"{category_hint}/{filename}" if category_hint else filename - # Upsert: if a previous experience exists for this workspace at a + # Upsert: if a previous workflow exists for this workspace at a # different path (e.g. user renamed the workspace), delete it after a # successful write so we keep one file per session. - existing = store.find_experience_by_workspace_id(workspace_id) + existing = store.find_workflow_by_workspace_id(workspace_id) try: - store.write("experiences", rel_path, md_content) + store.write("workflows", rel_path, md_content) except ValueError as exc: raise AppError(ErrorCode.INVALID_REQUEST, str(exc)) from exc if existing and existing.get("path") and existing["path"] != rel_path: try: - store.delete("experiences", existing["path"]) + store.delete("workflows", existing["path"]) except Exception: logger.warning( - "Failed to delete stale session experience at %s", + "Failed to delete stale session workflow at %s", existing.get("path"), exc_info=True, ) - return json_ok({"path": rel_path, "category": "experiences"}) + return json_ok({"path": rel_path, "category": "workflows"}) # ── helpers for session-scoped distillation ─────────────────────────────── @@ -269,16 +271,21 @@ def distill_experience(): def _apply_session_front_matter( content: str, workspace_id: str, workspace_name: str, -) -> str: - """Override / inject session-binding fields in the experience front matter. - - - Composes the visible ``title`` as ``Experience from : `` - using the LLM-emitted ``subtitle`` (preferred) or pre-existing - ``title``. The original ``subtitle`` field is removed from the - front matter once consumed. +) -> tuple[str, str, str]: + """Override / inject session-binding fields in the workflow front matter. + + - Sets the visible ``title`` to the agent-emitted descriptive + ``subtitle`` (preferred) or the pre-existing ``title``, with any + legacy ``Workflow from : `` prefix stripped. The ``subtitle`` + field is removed from the front matter once consumed. + - Consumes the agent-emitted short ``filename`` hint (removed from the + front matter) and returns it so the caller can name the file without + using the long descriptive title. - Stamps ``source_workspace_id`` and ``source_workspace_name`` so the file can be looked up on subsequent distillations. - Forces ``source: distill`` (idempotent if already set). + + Returns ``(content_with_front_matter, title_core, filename_hint)``. """ from data_formulator.knowledge.store import parse_front_matter @@ -287,27 +294,31 @@ def _apply_session_front_matter( meta = {} subtitle = str(meta.pop("subtitle", "") or "").strip() + filename_hint = str(meta.pop("filename", "") or "").strip() existing_title = str(meta.get("title", "") or "").strip() - # Strip any "Experience from : " prefix from a prior pass so - # update-mode runs don't double-prefix when the LLM echoes the title. - title_core = subtitle or _strip_experience_prefix(existing_title) + # Strip any legacy "Workflow from : " (or "Experience from") + # prefix so update-mode runs don't carry it forward. + title_core = subtitle or _strip_workflow_prefix(existing_title) if not title_core: title_core = workspace_name - new_title = f"Experience from {workspace_name}: {title_core}" - meta["title"] = new_title + meta["title"] = title_core meta["source"] = "distill" meta["source_workspace_id"] = workspace_id meta["source_workspace_name"] = workspace_name - return _serialize_front_matter(meta, body) + return _serialize_front_matter(meta, body), title_core, filename_hint + +_EXP_PREFIX_RE = re.compile(r"^\s*(?:Workflow|Experience) from .+?:\s*", re.IGNORECASE) -_EXP_PREFIX_RE = re.compile(r"^\s*Experience from .+?:\s*", re.IGNORECASE) +# Path separators, Windows-reserved chars and control chars that must never +# appear in a filename derived from untrusted LLM output. +_UNSAFE_FILENAME_CHARS = re.compile(r'[\\/:*?"<>|\x00-\x1f]+') -def _strip_experience_prefix(title: str) -> str: +def _strip_workflow_prefix(title: str) -> str: return _EXP_PREFIX_RE.sub("", title).strip() @@ -323,16 +334,22 @@ def _serialize_front_matter(meta: dict, body: str) -> str: return f"---\n{yaml_text}\n---\n\n{body_text}" -def _experience_filename(workspace_name: str) -> str: - """Derive a deterministic filename from the workspace name. +def _workflow_filename(title: str) -> str: + """Slugify an LLM-supplied name into a clean, safe ``.md`` filename. - Re-distilling the same session always lands on the same file. - Falls back to a literal slug when sanitisation rejects the name. + Re-distilling a session upserts by ``source_workspace_id`` (see caller), + so the file is replaced even when the name changes. ``safe_data_filename`` + enforces the security boundary (basename only, no ``.``/``..``); the slug + step just keeps separators and reserved chars out so the name is clean and + portable. Unicode (e.g. CJK) is preserved. """ from data_formulator.datalake.parquet_utils import safe_data_filename - slug = workspace_name.strip().replace(" ", "-").lower()[:80] or "session-experience" + cleaned = _UNSAFE_FILENAME_CHARS.sub("-", title) + cleaned = re.sub(r"\s+", "-", cleaned.strip()) + cleaned = re.sub(r"-{2,}", "-", cleaned) + slug = cleaned.strip(".-").lower()[:80] or "session-workflow" try: return safe_data_filename(f"{slug}.md") except ValueError: - return "session-experience.md" + return "session-workflow.md" diff --git a/src/api/knowledgeApi.ts b/src/api/knowledgeApi.ts index 7c149f3a..a722c00f 100644 --- a/src/api/knowledgeApi.ts +++ b/src/api/knowledgeApi.ts @@ -2,7 +2,7 @@ // Licensed under the MIT License. /** - * Knowledge API client — CRUD, search, and experience distillation. + * Knowledge API client — CRUD, search, and workflow distillation. * * All endpoints use POST with JSON body. Requests go through * {@link fetchWithIdentity} for identity headers and 401 retry. @@ -14,11 +14,10 @@ import { apiRequest } from '../app/apiClient'; // ── Types ──────────────────────────────────────────────────────────────── -export type KnowledgeCategory = 'rules' | 'experiences'; +export type KnowledgeCategory = 'rules' | 'workflows'; export interface KnowledgeItem { title: string; - tags: string[]; path: string; source: string; created: string; @@ -27,25 +26,24 @@ export interface KnowledgeItem { /** Rules only: if true the rule is always injected into the agent prompt. */ alwaysApply?: boolean; /** - * Experiences only: workspace id this experience was distilled from. + * Workflows only: workspace id this workflow was distilled from. * Set by the session-scoped distillation flow (design-docs/24); used - * by the KnowledgePanel to find the existing session experience. + * by the KnowledgePanel to find the existing session workflow. */ sourceWorkspaceId?: string; - /** Experiences only: workspace display name at distillation time. */ + /** Workflows only: workspace display name at distillation time. */ sourceWorkspaceName?: string; } export interface KnowledgeLimits { rule_description_max: number; rules: number; - experiences: number; + workflows: number; } export interface KnowledgeSearchResult { category: KnowledgeCategory; title: string; - tags: string[]; path: string; snippet: string; source: string; @@ -122,7 +120,7 @@ export async function searchKnowledge( return data.results ?? []; } -export interface DistillExperienceResult { +export interface DistillWorkflowResult { path: string; category: string; } @@ -134,7 +132,7 @@ export interface DistillExperienceResult { * a deterministic filename + title. `threads` carries one chronological * `events` list per leaf table on screen. */ -export interface SessionExperienceContext { +export interface SessionWorkflowContext { context_id?: string; workspace_id: string; workspace_name: string; @@ -146,18 +144,18 @@ export interface SessionExperienceContext { payload_notes?: string[]; } -export async function distillSessionExperience( - sessionContext: SessionExperienceContext, +export async function distillSessionWorkflow( + sessionContext: SessionWorkflowContext, model: Record, instruction?: string, timeoutSeconds?: number, signal?: AbortSignal, -): Promise { - const { data } = await apiRequest<{ path: string; category: string }>('/api/knowledge/distill-experience', { +): Promise { + const { data } = await apiRequest<{ path: string; category: string }>('/api/knowledge/distill-workflow', { method: 'POST', headers: JSON_HEADERS, body: JSON.stringify({ - experience_context: sessionContext, + workflow_context: sessionContext, model, user_instruction: instruction, timeout_seconds: timeoutSeconds, diff --git a/src/app/useKnowledgeStore.ts b/src/app/useKnowledgeStore.ts index 0a6ea65d..6adeb60c 100644 --- a/src/app/useKnowledgeStore.ts +++ b/src/app/useKnowledgeStore.ts @@ -5,7 +5,7 @@ * Knowledge state management — React hooks for knowledge CRUD & search. * * Uses plain React state (not Redux) because knowledge data is server-side - * and only needed by the KnowledgePanel and save-as-experience flows. + * and only needed by the KnowledgePanel and save-as-workflow flows. * Errors are dispatched to the global MessageSnackbar via dfActions.addMessages. */ @@ -40,16 +40,16 @@ export function useKnowledgeStore() { const { t } = useTranslation(); const [rules, setRules] = useState({ ...EMPTY_CATEGORY }); - const [experiences, setExperiences] = useState({ ...EMPTY_CATEGORY }); + const [workflows, setWorkflows] = useState({ ...EMPTY_CATEGORY }); const [searchResults, setSearchResults] = useState([]); const [searching, setSearching] = useState(false); - const DEFAULT_LIMITS: KnowledgeLimits = { rule_description_max: 100, rules: 350, experiences: 2000 }; + const DEFAULT_LIMITS: KnowledgeLimits = { rule_description_max: 100, rules: 350, workflows: 2000 }; const [limits, setLimits] = useState(DEFAULT_LIMITS); - const stateMap = { rules, experiences }; - const setterMap = useRef({ rules: setRules, experiences: setExperiences }); + const stateMap = { rules, workflows }; + const setterMap = useRef({ rules: setRules, workflows: setWorkflows }); const fetchList = useCallback(async (category: KnowledgeCategory) => { const setter = setterMap.current[category]; @@ -71,7 +71,7 @@ export function useKnowledgeStore() { const fetchAll = useCallback(async () => { await Promise.all([ fetchList('rules'), - fetchList('experiences'), + fetchList('workflows'), fetchKnowledgeLimits().then(setLimits).catch(() => { /* best-effort */ }), ]); }, [fetchList]); @@ -184,7 +184,7 @@ export function useKnowledgeStore() { return { rules, - experiences, + workflows, stateMap, limits, searchResults, diff --git a/src/i18n/locales/en/common.json b/src/i18n/locales/en/common.json index 7a52125b..8ef952df 100644 --- a/src/i18n/locales/en/common.json +++ b/src/i18n/locales/en/common.json @@ -875,9 +875,9 @@ "knowledge": { "title": "Agent Knowledge", "rules": "Rules", - "experiences": "Experiences", + "workflows": "Workflows", "rulesDescription": "Constraints and standards that agents must follow", - "experiencesDescription": "Reusable methods, tips, and knowledge distilled from analyses", + "workflowsDescription": "Reusable analysis workflows distilled from past sessions that agents can save and replay", "newItem": "New", "search": "Search", "searchPlaceholder": "Search knowledge...", @@ -902,29 +902,29 @@ "failedToSave": "Failed to save knowledge", "failedToDelete": "Failed to delete knowledge", "failedToSearch": "Search failed", - "saveAsExperience": "Save as Experience", - "saveAsExperienceTitle": "Save as Experience", - "distillHint": "Distill experience from this analysis for agents to reuse in future analysis.", + "saveAsExperience": "Save as Workflow", + "saveAsExperienceTitle": "Save as Workflow", + "distillHint": "Distill a workflow from this analysis for agents to save and replay in future sessions.", "distillFromHeading": "Distill from", "distillFromCaption": "Threads below will be sent to the LLM. Click a thread to inspect its events.", - "distillingOverlay": "Distilling experience… this may take a moment.", + "distillingOverlay": "Distilling workflow… this may take a moment.", "userInstruction": "User instruction (optional)", "userInstructionPlaceholder": "what to focus on, what to skip…", "distillationInstructions": "Distillation instructions (optional)", "distillationInstructionsPlaceholder": "e.g. focus on the data cleaning steps; skip exploratory chart variations; emphasise pitfalls we hit when joining tables…", - "distillExperience": "Distill Experience", - "distillStarted": "Distilling experience...", - "distilling": "Distilling experience...", - "distilled": "Experience saved", + "distillWorkflow": "Distill Workflow", + "distillStarted": "Distilling workflow...", + "distilling": "Distilling workflow...", + "distilled": "Workflow saved", "distillFailedRetry": "Save failed, retry", - "failedToDistill": "Failed to distill experience", - "distillSessionTitle": "Distill Session Experience", - "updateSessionTitle": "Update Session Experience", - "distillSessionHint": "Distill lessons from this analysis into a reusable knowledge document.", - "distillSessionUpdateHint": "Re-distill lessons from this analysis into the existing knowledge document.", + "failedToDistill": "Failed to distill workflow", + "distillSessionTitle": "Distill Session Workflow", + "updateSessionTitle": "Update Session Workflow", + "distillSessionHint": "Distill this analysis into a reusable workflow document that agents can replay.", + "distillSessionUpdateHint": "Re-distill this analysis into the existing workflow document.", "distillSessionNothing": "No completed analysis threads in this session yet.", "distillFromSession": "Distill from this session", - "experiencePlaceholderHint": "Save lessons learned", + "workflowPlaceholderHint": "Save this analysis as a workflow", "updateFromSession": "Update from this session", "updateFromSessionHint": "Refresh with new lessons", "addNewRule": "Add new rule", @@ -937,15 +937,21 @@ "itemCount": "({{count}})", "collapse": "Collapse", "expand": "Expand", - "emptyState": "Add rules or experiences to help AI agents work better.", - "rulesHint": "Rules — constraints the agent always follows. Click + to add your own.", - "experiencesHint": "Experiences — lessons distilled from your past analyses. Click the placeholder below to distill one from this session.", + "emptyState": "Add rules or workflows to help AI agents work better.", + "rulesHint": "Constraints the agent always follows.", + "workflowsHint": "Analyses distilled from past sessions that the agent can save and replay.", "markdownEditor": "Markdown Editor", "description": "Description", "descriptionPlaceholder": "Short summary of this rule (max {{max}} chars)", "alwaysApply": "Always loaded into AI", "alwaysApplyHint": "When enabled, this rule is always injected into every AI agent prompt, regardless of context", "charCount": "{{current}} / {{max}}", - "charCountExceeded": "Exceeds {{max}} character limit ({{current}} / {{max}})" + "charCountExceeded": "Exceeds {{max}} character limit ({{current}} / {{max}})", + "replay": "Replay", + "replayTooltip": "Replay this analysis on the current data", + "replayBusy": "The agent is busy — wait for it to finish before replaying.", + "replayNoData": "Load and focus a dataset before replaying a workflow.", + "replayStarted": "Replaying workflow on the current data…", + "replayPrompt": "Reproduce the following analysis workflow on the currently loaded data. Follow the steps in order, adapting any column references to the columns available in the current dataset. It's fine if the result isn't identical — reproduce the same overall analysis.\n\nBefore making large assumptions, check whether the current data can actually support the workflow. If there is a major discrepancy — e.g. a required field or measure is missing, the granularity or shape is very different, or a step has no sensible equivalent on this data — pause and ask me to confirm how to proceed (or briefly explain the mismatch and your proposed adaptation) instead of guessing. Minor differences (renamed columns, extra columns) can be adapted silently.\n\n{{content}}" } } diff --git a/src/i18n/locales/zh/common.json b/src/i18n/locales/zh/common.json index 1740e92a..244f93da 100644 --- a/src/i18n/locales/zh/common.json +++ b/src/i18n/locales/zh/common.json @@ -875,9 +875,9 @@ "knowledge": { "title": "Agent 知识", "rules": "规则", - "experiences": "经验", + "workflows": "工作流", "rulesDescription": "Agent 必须遵守的约束和编码规范", - "experiencesDescription": "从分析中提炼的可复用方法和技巧", + "workflowsDescription": "从过往会话中提炼、可供 Agent 保存与重放的可复用分析工作流", "newItem": "新建", "search": "搜索", "searchPlaceholder": "搜索知识...", @@ -902,29 +902,29 @@ "failedToSave": "保存知识失败", "failedToDelete": "删除知识失败", "failedToSearch": "搜索失败", - "saveAsExperience": "保存为经验", - "saveAsExperienceTitle": "保存为经验", - "distillHint": "从本次分析中提炼经验，供 Agent 在后续分析中复用。", + "saveAsExperience": "保存为工作流", + "saveAsExperienceTitle": "保存为工作流", + "distillHint": "从本次分析中提炼工作流，供 Agent 在后续会话中保存与重放。", "distillFromHeading": "提炼来源", "distillFromCaption": "以下线索将发送给 LLM。点击线索可查看其事件。", - "distillingOverlay": "正在提炼经验…请稍候。", + "distillingOverlay": "正在提炼工作流…请稍候。", "userInstruction": "用户指令（可选）", "userInstructionPlaceholder": "重点关注什么、跳过什么…", "distillationInstructions": "提炼指令（可选）", "distillationInstructionsPlaceholder": "例如：重点关注数据清洗步骤；跳过探索性图表变体；着重记录表连接时遇到的陷阱…", - "distillExperience": "提炼经验", - "distillStarted": "正在提炼经验...", - "distilling": "正在提炼经验...", - "distilled": "经验已保存", - "distillFailedRetry": "保存经验失败，重试", - "failedToDistill": "提炼经验失败", - "distillSessionTitle": "提炼会话经验", - "updateSessionTitle": "更新会话经验", - "distillSessionHint": "从本次分析中提炼经验，生成一篇可复用的知识文档。", - "distillSessionUpdateHint": "重新提炼本次分析的经验，覆盖现有的知识文档。", + "distillWorkflow": "提炼工作流", + "distillStarted": "正在提炼工作流...", + "distilling": "正在提炼工作流...", + "distilled": "工作流已保存", + "distillFailedRetry": "保存工作流失败，重试", + "failedToDistill": "提炼工作流失败", + "distillSessionTitle": "提炼会话工作流", + "updateSessionTitle": "更新会话工作流", + "distillSessionHint": "将本次分析提炼为一篇可供 Agent 重放的可复用工作流文档。", + "distillSessionUpdateHint": "将本次分析重新提炼到现有的工作流文档中。", "distillSessionNothing": "本会话还没有可提炼的分析线索。", "distillFromSession": "从本会话提炼", - "experiencePlaceholderHint": "保存分析中的经验", + "workflowPlaceholderHint": "将本次分析保存为工作流", "updateFromSession": "从本会话更新", "updateFromSessionHint": "用新经验刷新该条目", "addNewRule": "添加新规则", @@ -937,15 +937,21 @@ "itemCount": "（{{count}}）", "collapse": "收起", "expand": "展开", - "emptyState": "添加规则、技能或经验，帮助 AI Agent 更好地工作。", - "rulesHint": "规则 — Agent 始终遵守的约束。点击 + 添加你自己的规则。", - "experiencesHint": "经验 — 从你过往分析中提炼出的经验。点击下方占位项可从本会话提炼一条。", + "emptyState": "添加规则或工作流，帮助 AI Agent 更好地工作。", + "rulesHint": "Agent 始终遵守的约束。", + "workflowsHint": "从过往会话中提炼、Agent 可保存与重放的分析。", "markdownEditor": "Markdown 编辑器", "description": "描述", "descriptionPlaceholder": "规则的简短描述（最多 {{max}} 字符）", "alwaysApply": "始终加载到 AI", "alwaysApplyHint": "启用后，无论什么场景，此规则都会自动注入到每次 AI Agent 的提示词中", "charCount": "{{current}} / {{max}}", - "charCountExceeded": "超出 {{max}} 字符限制（{{current}} / {{max}}）" + "charCountExceeded": "超过 {{max}} 字符限制（{{current}} / {{max}}）", + "replay": "重放", + "replayTooltip": "在当前数据上重放此分析", + "replayBusy": "Agent 正忙——请等待其完成后再重放。", + "replayNoData": "请先加载并聚焦一个数据集，再重放工作流。", + "replayStarted": "正在当前数据上重放工作流…", + "replayPrompt": "在当前已加载的数据上复现以下分析流程。按顺序执行各步骤，并将其中的列引用调整为当前数据集中可用的列。结果不必完全一致——复现同样的整体分析即可。\n\n在做出较大假设之前，请先确认当前数据是否真的能支撑该流程。如果存在重大差异——例如缺少必需的字段或度量、数据粒度或结构差异很大、或某个步骤在当前数据上没有合理的对应方式——请暂停并向我确认如何继续（或简要说明不匹配之处及你建议的调整方案），而不要凭空猜测。对于细微差异（列被重命名、存在额外的列）可以直接静默调整。\n\n{{content}}" } } diff --git a/src/views/DataFrameTable.tsx b/src/views/DataFrameTable.tsx index dd32ecc9..afb03bbd 100644 --- a/src/views/DataFrameTable.tsx +++ b/src/views/DataFrameTable.tsx @@ -127,17 +127,24 @@ export const DataFrameTable: React.FC = ({ )} {displayCols.map((col, i) => { const desc = col !== '\u2026' ? columnDescriptions?.[col] : undefined; + if (desc) { + return ( + + + {col} + + + ); + } return ( - - - {col} - - + + {col} + ); })} diff --git a/src/views/DataSourceSidebar.tsx b/src/views/DataSourceSidebar.tsx index 7381a423..c0502fd0 100644 --- a/src/views/DataSourceSidebar.tsx +++ b/src/views/DataSourceSidebar.tsx @@ -157,7 +157,7 @@ export const DataSourceSidebar: React.FC<{ // appears when they try to add a new connector or link a folder. const [initialTab, setInitialTab] = useState<'sources' | 'sessions' | 'knowledge'>('sources'); - // External callers (e.g. SaveExperienceButton on success) can ask the + // External callers (e.g. workflow distill on success) can ask the // sidebar to open and switch to a specific tab. useEffect(() => { const handler = (e: Event) => { diff --git a/src/views/DataThread.tsx b/src/views/DataThread.tsx index 940cb4ef..248dbe75 100644 --- a/src/views/DataThread.tsx +++ b/src/views/DataThread.tsx @@ -1333,17 +1333,6 @@ let SingleThreadGroupView: FC<{ const mergeIds = derivedTable?.derive?.source as string[] | undefined; if (entry.role === 'instruction' && mergeNames && mergeNames.length > 0 && mergeIds && mergeIds.length > 0) { const nextKey = sourceSetKey(mergeIds); - // eslint-disable-next-line no-console - console.log('[merge-node check]', { - tableId, - parentTableId: parentTable?.id, - initialSourceIds, - prevSourceKey, - mergeIds, - mergeNames, - nextKey, - fires: nextKey !== prevSourceKey, - }); if (nextKey !== prevSourceKey) { const mergeColor = highlighted ? theme.palette.primary.main : theme.palette.text.secondary; timelineItems.push({ diff --git a/src/views/InteractionEntryCard.tsx b/src/views/InteractionEntryCard.tsx index 8cfd0000..79686ca2 100644 --- a/src/views/InteractionEntryCard.tsx +++ b/src/views/InteractionEntryCard.tsx @@ -242,6 +242,11 @@ export const InteractionEntryCard: React.FC = memo(({ // so they should read stronger than the agent's bubbles. backgroundColor: palette.bgcolor, border: `1px solid ${borderColor.component}`, + // Cap very long instructions (e.g. a replayed workflow) so the + // card stays compact; the full text scrolls within the cap. + maxHeight: 160, + overflowY: 'auto', + overscrollBehavior: 'contain', ...(highlighted ? { borderLeft: `2px solid ${palette.main}` } : {}), ...clickSx, }}> diff --git a/src/views/KnowledgePanel.tsx b/src/views/KnowledgePanel.tsx index 8d8a111f..05343a0a 100644 --- a/src/views/KnowledgePanel.tsx +++ b/src/views/KnowledgePanel.tsx @@ -4,16 +4,16 @@ /** * KnowledgePanel — panel for browsing and editing knowledge items. * - * Shows two collapsible sections: Rules (flat) and Experiences (flat). + * Shows two collapsible sections: Rules (flat) and Workflows (flat). * Items are tagged for organization; no subdirectory grouping. * Supports search, edit, and delete. Rules can be created directly by - * the user via the "+" affordance; experiences are produced by the + * the user via the "+" affordance; workflows are produced by the * agent's distillation flow (see SessionDistill). */ -import React, { useState, useCallback, useEffect, useRef } from 'react'; +import React, { useState, useCallback, useEffect } from 'react'; import { useTranslation } from 'react-i18next'; -import { useSelector } from 'react-redux'; +import { useSelector, useDispatch } from 'react-redux'; import { Box, Typography, @@ -26,7 +26,6 @@ import { DialogContent, DialogActions, CircularProgress, - Chip, Divider, } from '@mui/material'; import { alpha } from '@mui/material/styles'; @@ -34,6 +33,7 @@ import AddIcon from '@mui/icons-material/Add'; import DeleteOutlineIcon from '@mui/icons-material/DeleteOutline'; import DescriptionOutlinedIcon from '@mui/icons-material/DescriptionOutlined'; import SmartToyOutlinedIcon from '@mui/icons-material/SmartToyOutlined'; +import PlayArrowIcon from '@mui/icons-material/PlayArrow'; import RefreshIcon from '@mui/icons-material/Refresh'; import Editor from 'react-simple-code-editor'; @@ -41,9 +41,9 @@ import { useKnowledgeStore } from '../app/useKnowledgeStore'; import { deleteKnowledge, type KnowledgeCategory } from '../api/knowledgeApi'; import type { KnowledgeItem } from '../api/knowledgeApi'; import { borderColor, radius } from '../app/tokens'; -import { type DataFormulatorState } from '../app/dfSlice'; -import { isLeafDerivedTable, buildLeafEvents } from './experienceContext'; -import { SessionDistillDialog, findSessionExperience } from './SessionDistill'; +import { dfActions, type DataFormulatorState } from '../app/dfSlice'; +import { isLeafDerivedTable, buildLeafEvents } from './workflowContext'; +import { SessionDistillDialog, findSessionWorkflow } from './SessionDistill'; // Default file name and seed body for a brand-new rule. Rules are plain // Markdown — the user just edits the body; no front matter is required. @@ -58,19 +58,18 @@ Describe the constraints or conventions the agent should follow. interface ActionRowProps { icon: React.ReactNode; label: string; - hint: string; onClick: () => void; } -const ActionRow: React.FC = ({ icon, label, hint, onClick }) => ( +const ActionRow: React.FC = ({ icon, label, onClick }) => ( `1px solid ${alpha(theme.palette.primary.main, 0.5)}`, @@ -88,22 +87,12 @@ const ActionRow: React.FC = ({ icon, label, hint, onClick }) => userSelect: 'none', }} > - {icon} - - - {label} - - - {hint} - - + {icon} + + {label} + ); @@ -112,8 +101,9 @@ const ActionRow: React.FC = ({ icon, label, hint, onClick }) => export const KnowledgePanel: React.FC = () => { const { t } = useTranslation(); const store = useKnowledgeStore(); + const dispatch = useDispatch(); - // For the "distill from this session" placeholder under EXPERIENCES. + // For the "distill from this session" placeholder under WORKFLOWS. const tables = useSelector((s: DataFormulatorState) => s.tables); const charts = useSelector((s: DataFormulatorState) => s.charts); const conceptShelfItems = useSelector((s: DataFormulatorState) => s.conceptShelfItems); @@ -183,35 +173,6 @@ export const KnowledgePanel: React.FC = () => { setEditorLoading(false); }, [store]); - // Pending request to auto-open an entry once it appears in the store - // (e.g. after the SessionDistillDialog finishes distilling). - const pendingOpenRef = useRef<{ category: KnowledgeCategory; path: string } | null>(null); - - useEffect(() => { - const handler = (e: Event) => { - const detail = (e as CustomEvent).detail || {}; - const category = (detail.category as KnowledgeCategory | undefined) ?? 'experiences'; - const path = detail.path as string | undefined; - if (path) { - pendingOpenRef.current = { category, path }; - } - }; - window.addEventListener('open-knowledge-panel', handler); - return () => window.removeEventListener('open-knowledge-panel', handler); - }, []); - - // When the requested entry shows up in the store, open its editor. - useEffect(() => { - const pending = pendingOpenRef.current; - if (!pending) return; - const cat = store.stateMap[pending.category]; - if (!cat?.loaded) return; - const item = cat.items.find(i => i.path === pending.path); - if (!item) return; - pendingOpenRef.current = null; - openEditDialog(pending.category, item); - }, [store.stateMap, openEditDialog]); - const handleSave = useCallback(async () => { if (!editorPath.trim() || !editorContent.trim()) return; setEditorSaving(true); @@ -237,9 +198,9 @@ export const KnowledgePanel: React.FC = () => { }, [deleteTarget, store]); // ── Distill from current session ──────────────────────────────────── - // The EXPERIENCES placeholder under EXPERIENCES is bound to the + // The WORKFLOWS placeholder is bound to the // active workspace. When the workspace already has a distilled - // experience (matched by `sourceWorkspaceId` in front matter) we + // workflow (matched by `sourceWorkspaceId` in front matter) we // expose an inline ⟳ Update affordance on the existing entry; // otherwise the placeholder opens the dialog in *create* mode. // See design-docs/24-session-scoped-distillation.md. @@ -265,9 +226,9 @@ export const KnowledgePanel: React.FC = () => { const selectedModel = allModels.find(m => m.id === selectedModelId); const canDistillFromSession = hasDistillableSession && !!selectedModel && !!activeWorkspace; - const sessionExperience = React.useMemo( - () => findSessionExperience( - store.stateMap['experiences'].items, + const sessionWorkflow = React.useMemo( + () => findSessionWorkflow( + store.stateMap['workflows'].items, activeWorkspace?.id, ), [store.stateMap, activeWorkspace?.id], @@ -278,6 +239,21 @@ export const KnowledgePanel: React.FC = () => { setSessionDialogOpen(true); }, []); + // ── Replay a workflow ──────────────────────────────────────────── + // Reads the workflow body and asks the data agent (in SimpleChartRecBox) + // to reproduce the captured workflow on the currently loaded data. v1 is + // deliberately simple: we hand the whole workflow to the agent in one + // request via a window event and let it figure out the rest. + // See discussion/replayable-experience-workflow.md. + const handleReplay = useCallback(async (item: KnowledgeItem) => { + const content = await store.read('workflows', item.path); + if (content == null) return; + const prompt = t('knowledge.replayPrompt', { content }); + window.dispatchEvent(new CustomEvent('df-replay-workflow', { + detail: { prompt, title: item.title }, + })); + }, [store, t]); + // ── Render section ────────────────────────────────────────────────── @@ -285,81 +261,85 @@ export const KnowledgePanel: React.FC = () => { category: KnowledgeCategory, item: KnowledgeItem, ) => { - const displayName = item.path || item.title; + const displayTitle = (item.title || '').replace(/^\s*(?:Workflow|Experience) from .+?:\s*/i, '').trim(); + const primary = displayTitle || item.title || item.path; return ( openEditDialog(category, item)} sx={{ display: 'flex', alignItems: 'flex-start', gap: 0.75, - px: 1.5, py: 0.75, + px: 1.5, py: 0.625, cursor: 'pointer', color: 'text.primary', '&:hover': { bgcolor: 'action.hover' }, - '&:hover .item-actions': { visibility: 'visible' }, + '&:hover .item-actions': { display: 'inline-flex' }, userSelect: 'none', }} > - + - - {displayName} + + {primary} - {item.tags.length > 0 && ( - - {item.tags.map(tag => ( - - ))} - - )} {item.source === 'agent_summarized' && ( )} - + + {category === 'workflows' && ( + + { e.stopPropagation(); handleReplay(item); }} + sx={{ + p: 0.25, + color: 'primary.main', + '&:hover': { bgcolor: theme => alpha(theme.palette.primary.main, 0.08) }, + }} + > + + + + )} { e.stopPropagation(); setDeleteTarget({ category, path: item.path, title: item.title }); }} - sx={{ p: 0.25, color: 'text.secondary', '&:hover': { color: 'error.main' } }} + sx={{ p: 0.25, display: 'none', color: 'text.secondary', '&:hover': { color: 'error.main' } }} > - + ); - }, [openEditDialog, t]); + }, [openEditDialog, t, handleReplay]); const renderCategorySection = useCallback(( category: KnowledgeCategory, label: string, + hint: string, ) => { const state = store.stateMap[category]; // Persistent action row at the top of the section. Rules: opens - // the create dialog. Experiences: opens the session distill + // the create dialog. Workflows: opens the session distill // dialog in create or update mode depending on whether the active - // workspace already has a distilled experience. + // workspace already has a distilled workflow. // See design-docs/24-session-scoped-distillation.md. const renderActionRow = () => { if (category === 'rules') { return ( } + icon={} label={t('knowledge.addNewRule', { defaultValue: 'Add new rule' })} - hint={t('knowledge.addNewRuleHint', { defaultValue: 'Set a convention for the agent' })} onClick={() => openCreateDialog('rules')} /> ); } - // experiences + // workflows if (!canDistillFromSession) { // No active workspace, no model, or no distillable thread // yet — show a passive hint instead of a dead action. @@ -370,15 +350,12 @@ export const KnowledgePanel: React.FC = () => { ); } - const updateMode = !!sessionExperience; + const updateMode = !!sessionWorkflow; if (sessionDistilling) { return ( } - label={t('knowledge.distilling', { defaultValue: 'Distilling experience…' })} - hint={updateMode - ? t('knowledge.updateFromSessionHint', { defaultValue: 'Refresh with new lessons' }) - : t('knowledge.experiencePlaceholderHint', { defaultValue: 'Save lessons learned' })} + icon={} + label={t('knowledge.distilling', { defaultValue: 'Distilling workflow…' })} onClick={() => openSessionDistillDialog(updateMode)} /> ); @@ -386,32 +363,44 @@ export const KnowledgePanel: React.FC = () => { return ( - : } + ? + : } label={updateMode ? t('knowledge.updateFromSession', { defaultValue: 'Update from this session' }) : t('knowledge.distillFromSession', { defaultValue: 'Distill from this session' })} - hint={updateMode - ? t('knowledge.updateFromSessionHint', { defaultValue: 'Refresh with new lessons' }) - : t('knowledge.experiencePlaceholderHint', { defaultValue: 'Save lessons learned' })} onClick={() => openSessionDistillDialog(updateMode)} /> ); }; return ( - + - + {label} + {/* Always-visible guidance for the section, set off by a + subtle left accent line below the title. */} + alpha(theme.palette.primary.main, 0.25), + }} + > + + {hint} + + + {state.loading && ( @@ -421,36 +410,18 @@ export const KnowledgePanel: React.FC = () => { {state.items.map(item => renderItem(category, item))} ); - }, [store.stateMap, renderItem, openCreateDialog, t, canDistillFromSession, sessionExperience, sessionDistilling, openSessionDistillDialog]); + }, [store.stateMap, renderItem, openCreateDialog, t, canDistillFromSession, sessionWorkflow, sessionDistilling, openSessionDistillDialog]); // ── Main render ───────────────────────────────────────────────────── return ( - {/* Persistent hint — explains Rules vs Experiences without - requiring the user to scroll past empty-state messages. */} - - - {t('knowledge.rulesHint')} - - - {t('knowledge.experiencesHint')} - - - - {/* Content area */} + {/* Content area. Rules vs Workflows guidance is surfaced via an + info icon next to each section title (see renderCategorySection). */} - {renderCategorySection('rules', t('knowledge.rules'))} - {renderCategorySection('experiences', t('knowledge.experiences'))} + {renderCategorySection('rules', t('knowledge.rules'), t('knowledge.rulesHint'))} + {renderCategorySection('workflows', t('knowledge.workflows'), t('knowledge.workflowsHint'))} @@ -515,22 +486,6 @@ export const KnowledgePanel: React.FC = () => { }} /> - {(() => { - const bodyLimit = store.limits[editorCategory as keyof typeof store.limits] as number | undefined; - if (!bodyLimit) return null; - const bodyLen = editorContent.trim().length; - const exceeded = bodyLen > bodyLimit; - return ( - bodyLimit * 0.9 ? 'warning.main' : 'text.disabled', - }}> - {exceeded - ? t('knowledge.charCountExceeded', { max: bodyLimit, current: bodyLen }) - : t('knowledge.charCount', { max: bodyLimit, current: bodyLen })} - - ); - })()} )} @@ -555,7 +510,6 @@ export const KnowledgePanel: React.FC = () => { editorSaving || !editorContent.trim() || !editorPath.trim() - || editorContent.trim().length > (store.limits[editorCategory as keyof typeof store.limits] as number ?? Infinity) } variant="contained" sx={{ textTransform: 'none', fontSize: 12 }} diff --git a/src/views/SessionDistill.tsx b/src/views/SessionDistill.tsx index fbff4f3b..fd9efeae 100644 --- a/src/views/SessionDistill.tsx +++ b/src/views/SessionDistill.tsx @@ -2,19 +2,19 @@ // Licensed under the MIT License. /** - * SessionDistill — session-scoped experience distillation. + * SessionDistill — session-scoped workflow distillation. * * Replaces the old per-result distillation flow with a single * session-bound entry. See design-docs/24-session-scoped-distillation.md. * * Exports: - * - buildSessionExperienceContext(workspace, threads): state-independent + * - buildSessionWorkflowContext(workspace, threads): state-independent * payload builder (with size budgeting, see §3.5 of the design doc). * - collectSessionThreads(tables, charts, fields): leaf discovery + per-leaf * event walk against live DataFormulator state. * - SessionDistillDialog: the dialog used by KnowledgePanel for both * create and update modes. - * - findSessionExperience: lookup an existing session experience by + * - findSessionWorkflow: lookup an existing session workflow by * workspace id. */ @@ -51,16 +51,16 @@ import { import { store, type AppDispatch } from '../app/store'; import { handleApiError } from '../app/errorHandler'; import { - distillSessionExperience, + distillSessionWorkflow, type KnowledgeItem, - type SessionExperienceContext, + type SessionWorkflowContext, } from '../api/knowledgeApi'; import { buildLeafEvents, buildDistillModelConfig, isLeafDerivedTable, TOOL_USES_CODE_FONT, -} from './experienceContext'; +} from './workflowContext'; // --------------------------------------------------------------------------- // Payload size budget (design-docs/24 §3.5) @@ -81,7 +81,7 @@ const SESSION_EVENT_BUDGET = 60_000; // bytes of JSON-serialized events // --------------------------------------------------------------------------- /** - * One pre-built thread, ready for `buildSessionExperienceContext`. + * One pre-built thread, ready for `buildSessionWorkflowContext`. * * Callers produce these by walking their own tables (see * `collectSessionThreads` for the in-app implementation) or with hand-built @@ -97,7 +97,7 @@ export interface SessionThread { export interface BuildSessionResult { /** Payload as it will be sent (after trimming). */ - payload: SessionExperienceContext; + payload: SessionWorkflowContext; /** Display threads with labels for the preview UI (post-trim). */ threads: SessionThread[]; /** Aggregate stats for the preview (post-trim). */ @@ -107,14 +107,14 @@ export interface BuildSessionResult { } // --------------------------------------------------------------------------- -// findSessionExperience +// findSessionWorkflow // --------------------------------------------------------------------------- /** - * Find the experience entry distilled from the given workspace, if any. + * Find the workflow entry distilled from the given workspace, if any. * Returns the first match; the backend ensures at most one per workspace. */ -export function findSessionExperience( +export function findSessionWorkflow( items: KnowledgeItem[] | undefined, workspaceId: string | undefined, ): KnowledgeItem | undefined { @@ -132,7 +132,7 @@ export function findSessionExperience( * * Threads with no user message are filtered out. Returns `[]` when the * session has no distillable thread. Not used in tests — tests construct - * `SessionThread[]` directly and call `buildSessionExperienceContext`. + * `SessionThread[]` directly and call `buildSessionWorkflowContext`. */ export function collectSessionThreads( tables: DictTable[], @@ -162,16 +162,16 @@ export function collectSessionThreads( } // --------------------------------------------------------------------------- -// buildSessionExperienceContext — pure (workspace, threads) → payload +// buildSessionWorkflowContext — pure (workspace, threads) → payload // --------------------------------------------------------------------------- /** - * Assemble the multi-thread payload sent to `/api/knowledge/distill-experience`. + * Assemble the multi-thread payload sent to `/api/knowledge/distill-workflow`. * * State-independent: takes pre-built threads and a workspace identity. * Returns `null` when `threads` is empty. */ -export function buildSessionExperienceContext( +export function buildSessionWorkflowContext( workspace: { id: string; displayName: string }, threads: SessionThread[], ): BuildSessionResult | null { @@ -179,7 +179,7 @@ export function buildSessionExperienceContext( const { trimmedThreads, notes } = trimToBudget(threads, SESSION_EVENT_BUDGET); - const payload: SessionExperienceContext = { + const payload: SessionWorkflowContext = { context_id: workspace.id, workspace_id: workspace.id, workspace_name: workspace.displayName, @@ -308,7 +308,7 @@ export const SessionDistillDialog: React.FC = ({ const built = useMemo(() => { if (!open || !activeWorkspace) return null; const threads = collectSessionThreads(tables, charts, conceptShelfItems); - return buildSessionExperienceContext(activeWorkspace, threads); + return buildSessionWorkflowContext(activeWorkspace, threads); }, [open, activeWorkspace, tables, charts, conceptShelfItems]); const [userInstruction, setUserInstruction] = useState(''); @@ -330,6 +330,10 @@ export const SessionDistillDialog: React.FC = ({ setStatus('running'); onRunningChange?.(true); const instruction = userInstruction.trim() || undefined; + // Close the dialog right away — distillation continues in the + // background and surfaces its result via the events/toast below. + setUserInstruction(''); + onClose(); try { const modelConfig = buildDistillModelConfig(selectedModel as ModelConfig); @@ -338,7 +342,7 @@ export const SessionDistillDialog: React.FC = ({ const timeoutId = setTimeout(() => controller.abort(), timeoutSeconds * 1000); let result; try { - result = await distillSessionExperience( + result = await distillSessionWorkflow( built.payload, modelConfig, instruction, timeoutSeconds, controller.signal, ); } finally { @@ -351,14 +355,12 @@ export const SessionDistillDialog: React.FC = ({ value: t('knowledge.distilled'), })); window.dispatchEvent(new CustomEvent('knowledge-changed', { - detail: { category: 'experiences' }, + detail: { category: 'workflows' }, })); window.dispatchEvent(new CustomEvent('open-knowledge-panel', { - detail: { category: 'experiences', path: result.path }, + detail: { category: 'workflows', path: result.path }, })); setStatus('idle'); - setUserInstruction(''); - onClose(); } catch (e: unknown) { setStatus('failed'); handleApiError(e, 'knowledge'); @@ -375,8 +377,8 @@ export const SessionDistillDialog: React.FC = ({ {updateMode - ? t('knowledge.updateSessionTitle', { defaultValue: 'Update Session Experience' }) - : t('knowledge.distillSessionTitle', { defaultValue: 'Distill Session Experience' })} + ? t('knowledge.updateSessionTitle', { defaultValue: 'Update Session Workflow' }) + : t('knowledge.distillSessionTitle', { defaultValue: 'Distill Session Workflow' })} = ({ {updateMode ? t('knowledge.distillSessionUpdateHint', { - defaultValue: 'Re-distill lessons from this analysis into the existing knowledge document.', + defaultValue: 'Re-distill this analysis into the existing workflow document.', }) : t('knowledge.distillSessionHint', { - defaultValue: 'Distill lessons from this analysis into a reusable knowledge document.', + defaultValue: 'Distill this analysis into a reusable workflow document that agents can replay.', })} @@ -479,7 +481,7 @@ export const SessionDistillDialog: React.FC = ({ ? t('knowledge.distilling') : updateMode ? t('knowledge.updateSession', { defaultValue: 'Update' }) - : t('knowledge.distillExperience')} + : t('knowledge.distillWorkflow')} diff --git a/src/views/SimpleChartRecBox.tsx b/src/views/SimpleChartRecBox.tsx index 23b30c7c..e94d2192 100644 --- a/src/views/SimpleChartRecBox.tsx +++ b/src/views/SimpleChartRecBox.tsx @@ -38,8 +38,6 @@ import AddIcon from '@mui/icons-material/Add'; import TipsAndUpdatesIcon from '@mui/icons-material/TipsAndUpdates'; import StopIcon from '@mui/icons-material/Stop'; -import AutoGraphIcon from '@mui/icons-material/AutoGraph'; -import DescriptionOutlinedIcon from '@mui/icons-material/DescriptionOutlined'; import InsertDriveFileOutlinedIcon from '@mui/icons-material/InsertDriveFileOutlined'; import { borderColor, transition } from '../app/tokens'; import { Theme } from '@mui/material/styles'; @@ -70,7 +68,7 @@ const AgentWorkingOverlay: FC<{ message?: string; elapsed?: number; theme: Theme }}> - + {t('chartRec.agentWorking')} @@ -96,13 +94,13 @@ const AgentWorkingOverlay: FC<{ message?: string; elapsed?: number; theme: Theme )} {latestMessage}{elapsedSuffix} @@ -1354,6 +1352,39 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ exploreFromChat(prompt, undefined, displayPrompt); }, [reportFromChat, exploreFromChat, selectedAgent, clarificationQuestions, clarifyAnswers]); + // Replay a workflow: the KnowledgePanel fires `df-replay-workflow` + // with a prompt describing the captured workflow; we hand it straight to + // the data agent on the currently focused dataset. v1 is deliberately + // simple — one request, let the agent reproduce the analysis on its own. + // See discussion/replayable-experience-workflow.md. + useEffect(() => { + const handler = (e: Event) => { + const prompt = (e as CustomEvent).detail?.prompt as string | undefined; + if (!prompt) return; + if (isChatFormulating) { + dispatch(dfActions.addMessages({ + timestamp: Date.now(), type: 'error', + component: 'data-agent', value: t('knowledge.replayBusy'), + })); + return; + } + if (!focusedTableId) { + dispatch(dfActions.addMessages({ + timestamp: Date.now(), type: 'error', + component: 'data-agent', value: t('knowledge.replayNoData'), + })); + return; + } + dispatch(dfActions.addMessages({ + timestamp: Date.now(), type: 'info', + component: 'data-agent', value: t('knowledge.replayStarted'), + })); + exploreFromChat(prompt); + }; + window.addEventListener('df-replay-workflow', handler); + return () => window.removeEventListener('df-replay-workflow', handler); + }, [exploreFromChat, isChatFormulating, focusedTableId, dispatch, t]); + const resumeFromClarification = useCallback((responses: ClarificationResponse[]) => { if (!pendingClarification) return; // Pass the formatted display string as `prompt` — it powers both the @@ -1744,9 +1775,6 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ '&:hover': { backgroundColor: alpha(isReportMode ? theme.palette.warning.main : theme.palette.primary.main, 0.08) }, }} > - {selectedAgent === 'explore' - ? - : } {selectedAgent === 'explore' ? t('chartRec.modeExplore') : t('chartRec.modeReport')} @@ -1761,7 +1789,7 @@ export const SimpleChartRecBox: FC<{ onInputFocus?: () => void }> = function ({ submitChat(t('chartRec.exploreIdeasPrompt'), undefined, t('chartRec.askedForRecommendations'))} > diff --git a/src/views/experienceContext.ts b/src/views/workflowContext.ts similarity index 98% rename from src/views/experienceContext.ts rename to src/views/workflowContext.ts index 98ec1c80..dac6e006 100644 --- a/src/views/experienceContext.ts +++ b/src/views/workflowContext.ts @@ -2,8 +2,8 @@ // Licensed under the MIT License. /** - * experienceContext — pure helpers that turn DataFormulator state into - * the timeline payload sent to `/api/knowledge/distill-experience`. + * workflowContext — pure helpers that turn DataFormulator state into + * the timeline payload sent to `/api/knowledge/distill-workflow`. * * No React, no Redux. Used by: * - SessionDistill.collectSessionThreads (live distillation) diff --git a/tests/backend/agents/test_agent_knowledge_integration.py b/tests/backend/agents/test_agent_knowledge_integration.py index 3efc65df..4d738635 100644 --- a/tests/backend/agents/test_agent_knowledge_integration.py +++ b/tests/backend/agents/test_agent_knowledge_integration.py @@ -62,7 +62,7 @@ def user_home(tmp_path): rules_dir.mkdir(parents=True) (rules_dir / "roi.md").write_text(RULE_MD, encoding="utf-8") - exp_dir = tmp_path / "knowledge" / "experiences" / "cleaning" + exp_dir = tmp_path / "knowledge" / "workflows" / "cleaning" exp_dir.mkdir(parents=True) (exp_dir / "missing.md").write_text(SKILL_MD, encoding="utf-8") @@ -170,11 +170,11 @@ def test_no_match_no_injection(self, mock_client, mock_workspace, user_home): def test_max_five_items(self, mock_client, mock_workspace, tmp_path): rules_dir = tmp_path / "knowledge" / "rules" rules_dir.mkdir(parents=True) - exp_dir = tmp_path / "knowledge" / "experiences" / "common" + exp_dir = tmp_path / "knowledge" / "workflows" / "common" exp_dir.mkdir(parents=True) for i in range(10): (exp_dir / f"exp-{i}.md").write_text( - f"---\ntitle: Common Experience {i}\ntags: [common]\n" + f"---\ntitle: Common Workflow {i}\ntags: [common]\n" f"created: 2026-04-26\nupdated: 2026-04-26\n---\n" f"Content about common topic {i}.\n", encoding="utf-8", @@ -247,7 +247,7 @@ def test_agent_works_without_knowledge(self, mock_client, mock_workspace): def test_empty_knowledge_dir(self, mock_client, mock_workspace, tmp_path): """Agent with empty knowledge dir works normally.""" (tmp_path / "knowledge" / "rules").mkdir(parents=True) - (tmp_path / "knowledge" / "experiences").mkdir(parents=True) + (tmp_path / "knowledge" / "workflows").mkdir(parents=True) agent = _make_agent(mock_client, mock_workspace, tmp_path) prompt = agent._build_system_prompt() assert "User Rules" not in prompt diff --git a/tests/backend/agents/test_experience_distill.py b/tests/backend/agents/test_workflow_distill.py similarity index 74% rename from tests/backend/agents/test_experience_distill.py rename to tests/backend/agents/test_workflow_distill.py index a3b823c8..e44d4a86 100644 --- a/tests/backend/agents/test_experience_distill.py +++ b/tests/backend/agents/test_workflow_distill.py @@ -1,13 +1,13 @@ # Copyright (c) Microsoft Corporation. # Licensed under the MIT License. -"""Tests for ExperienceDistillAgent and the /api/knowledge/distill-experience endpoint. +"""Tests for WorkflowDistillAgent and the /api/knowledge/distill-workflow endpoint. Covers: -- _extract_context_summary correctly extracts experience context +- _extract_context_summary correctly extracts workflow context - Output Markdown includes valid YAML front matter - front matter contains source: distill and source metadata -- Generated experience file written to correct directory +- Generated workflow file written to correct directory - category_hint controls sub-directory """ @@ -18,8 +18,14 @@ import flask import pytest -from data_formulator.agents.agent_experience_distill import ExperienceDistillAgent -from data_formulator.knowledge.store import parse_front_matter +from data_formulator.agents.agent_workflow_distill import WorkflowDistillAgent +from data_formulator.knowledge.store import ( + KNOWLEDGE_LIMITS, + WORKFLOW_HARD_MAX, + parse_front_matter, +) + +WORKFLOW_SOFT_LIMIT = KNOWLEDGE_LIMITS["workflows"] pytestmark = [pytest.mark.backend] @@ -73,7 +79,7 @@ }, ] -SAMPLE_EXPERIENCE_CONTEXT = { +SAMPLE_WORKFLOW_CONTEXT = { "context_id": "ws-1", "workspace_id": "ws-1", "workspace_name": "Sales Region Analysis", @@ -86,7 +92,7 @@ class TestExtractContextSummary: def test_renders_each_event_type(self): - summary = ExperienceDistillAgent._extract_context_summary(SAMPLE_EXPERIENCE_CONTEXT) + summary = WorkflowDistillAgent._extract_context_summary(SAMPLE_WORKFLOW_CONTEXT) # message events assert "[user→data-agent/prompt]" in summary assert "Show sales by region" in summary @@ -113,7 +119,7 @@ def test_renders_each_event_type(self): assert "encoding: x=region(nominal)" in summary def test_empty_events_returns_marker(self): - summary = ExperienceDistillAgent._extract_context_summary({}) + summary = WorkflowDistillAgent._extract_context_summary({}) assert summary == "(empty context)" def test_user_content_is_not_displaycontent(self): @@ -131,7 +137,7 @@ def test_user_content_is_not_displaycontent(self): }], }], } - summary = ExperienceDistillAgent._extract_context_summary(ctx) + summary = WorkflowDistillAgent._extract_context_summary(ctx) assert "raw text" in summary def test_skips_non_dict_events(self): @@ -140,7 +146,7 @@ def test_skips_non_dict_events(self): {"type": "message", "from": "user", "to": "data-agent", "role": "prompt", "content": "ok"}, ]}]} - summary = ExperienceDistillAgent._extract_context_summary(ctx) + summary = WorkflowDistillAgent._extract_context_summary(ctx) assert "[user→data-agent/prompt]" in summary # No crashes; the bogus entries are silently dropped. @@ -154,7 +160,7 @@ def test_create_table_basic(self): "sample_rows": [{"a": 1}], "code": "x = 1", }]}]} - summary = ExperienceDistillAgent._extract_context_summary(ctx) + summary = WorkflowDistillAgent._extract_context_summary(ctx) assert "[create_table] t1" in summary def test_create_chart_without_encoding(self): @@ -163,7 +169,7 @@ def test_create_chart_without_encoding(self): "related_table_id": "t1", "mark_or_type": "line", }]}]} - summary = ExperienceDistillAgent._extract_context_summary(ctx) + summary = WorkflowDistillAgent._extract_context_summary(ctx) assert "[create_chart] line on t1" in summary assert "encoding:" not in summary @@ -187,7 +193,7 @@ def test_renders_multi_thread_with_headers(self): }, ], } - summary = ExperienceDistillAgent._extract_context_summary(ctx) + summary = WorkflowDistillAgent._extract_context_summary(ctx) assert "### Thread 1 (id=leaf-a)" in summary assert "### Thread 2 (id=leaf-b)" in summary assert "load gas prices" in summary @@ -236,10 +242,10 @@ def _mock_client(self): def test_produces_valid_markdown(self): client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) with patch.object(agent, "_call_llm", return_value=MOCK_CONTEXT_RESPONSE): - result = agent.run(SAMPLE_EXPERIENCE_CONTEXT) + result = agent.run(SAMPLE_WORKFLOW_CONTEXT) assert result.startswith("---") meta, body = parse_front_matter(result) @@ -249,11 +255,11 @@ def test_produces_valid_markdown(self): def test_fallback_front_matter_added(self): client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) no_fm_response = "# Sales Analysis\n\nJust some content." with patch.object(agent, "_call_llm", return_value=no_fm_response): - result = agent.run(SAMPLE_EXPERIENCE_CONTEXT) + result = agent.run(SAMPLE_WORKFLOW_CONTEXT) assert result.startswith("---") meta, _ = parse_front_matter(result) @@ -261,11 +267,11 @@ def test_fallback_front_matter_added(self): assert meta["source_context"] == "ws-1" def test_retries_once_when_body_too_long(self): - """If first LLM call produces body > limit, agent retries with condensation prompt.""" + """If first LLM call produces body over the soft target, agent retries with condensation prompt.""" client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) - long_body = "x" * 3000 + long_body = "x" * (WORKFLOW_SOFT_LIMIT + 1000) long_response = ( "---\ntitle: Long\ntags: []\ncreated: 2026-01-01\n" "updated: 2026-01-01\nsource: distill\nsource_context: t1\n---\n\n" @@ -283,18 +289,18 @@ def fake_call_llm(messages): return short_response with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - result = agent.run(SAMPLE_EXPERIENCE_CONTEXT) + result = agent.run(SAMPLE_WORKFLOW_CONTEXT) assert call_count == 2 _, body = parse_front_matter(result) - assert len(body.strip()) <= 2000 + assert len(body.strip()) <= WORKFLOW_SOFT_LIMIT def test_retry_asks_for_slack_under_limit(self): - """The retry prompt asks the model for less than the hard limit.""" + """The retry prompt asks the model for less than the soft target.""" client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) - long_body = "x" * 3000 + long_body = "x" * (WORKFLOW_SOFT_LIMIT + 1000) long_response = ( "---\ntitle: L\ntags: []\ncreated: 2026-01-01\n" "updated: 2026-01-01\nsource: distill\nsource_context: t1\n---\n\n" @@ -309,21 +315,21 @@ def fake_call_llm(messages): return long_response if len(captured) == 1 else MOCK_CONTEXT_RESPONSE with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - agent.run(SAMPLE_EXPERIENCE_CONTEXT) + agent.run(SAMPLE_WORKFLOW_CONTEXT) assert len(captured) == 2 retry_prompt = captured[1][-1]["content"] - # Must mention the slacked target (limit minus margin), not the raw limit. - expected_target = 2000 - agent.RETRY_MARGIN - assert f"within {expected_target} characters" in retry_prompt + # Must mention the slacked target (soft limit minus margin). + expected_target = WORKFLOW_SOFT_LIMIT - agent.RETRY_MARGIN + assert f"around {expected_target} characters" in retry_prompt def test_hard_trims_when_retry_still_over_limit(self): - """If the retry still overshoots, body is hard-trimmed to fit the limit.""" + """If the retry still blows past the hard ceiling, body is hard-trimmed to fit it.""" client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) - first_body = "x" * 3000 - retry_body = "y" * 2014 # mimics the real-world failure: 14 over + first_body = "x" * (WORKFLOW_SOFT_LIMIT + 1000) + retry_body = "y" * (WORKFLOW_HARD_MAX + 14) # mimics retry still over the ceiling front_matter = ( "---\ntitle: T\ntags: []\ncreated: 2026-01-01\n" "updated: 2026-01-01\nsource: distill\nsource_context: t1\n---\n\n" @@ -339,13 +345,13 @@ def fake_call_llm(messages): return resp with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - result = agent.run(SAMPLE_EXPERIENCE_CONTEXT) + result = agent.run(SAMPLE_WORKFLOW_CONTEXT) # Both LLM calls happened. assert call_count == 2 - # Final body fits the hard limit (no save failure). + # Final body fits the hard ceiling (no save failure). _, body = parse_front_matter(result) - assert len(body.strip()) <= 2000 + assert len(body.strip()) <= WORKFLOW_HARD_MAX # Truncation marker is present so the user can see it was trimmed. assert "truncated" in body # Front matter preserved. @@ -355,7 +361,7 @@ def fake_call_llm(messages): def test_no_retry_when_body_within_limit(self): """If first LLM call is within limit, no retry happens.""" client = self._mock_client() - agent = ExperienceDistillAgent(client=client) + agent = WorkflowDistillAgent(client=client) call_count = 0 @@ -365,14 +371,14 @@ def fake_call_llm(messages): return MOCK_CONTEXT_RESPONSE with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - agent.run(SAMPLE_EXPERIENCE_CONTEXT) + agent.run(SAMPLE_WORKFLOW_CONTEXT) assert call_count == 1 def test_language_instruction_injected_into_system_prompt(self): client = self._mock_client() zh_instruction = "[LANGUAGE INSTRUCTION]\nWrite in Simplified Chinese." - agent = ExperienceDistillAgent(client=client, language_instruction=zh_instruction) + agent = WorkflowDistillAgent(client=client, language_instruction=zh_instruction) captured_messages = [] @@ -381,7 +387,7 @@ def fake_call_llm(messages): return MOCK_CONTEXT_RESPONSE with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - agent.run(SAMPLE_EXPERIENCE_CONTEXT) + agent.run(SAMPLE_WORKFLOW_CONTEXT) system_content = captured_messages[0]["content"] assert "[LANGUAGE INSTRUCTION]" in system_content @@ -389,7 +395,7 @@ def fake_call_llm(messages): def test_language_code_zh_injects_chinese_instruction(self): client = self._mock_client() - agent = ExperienceDistillAgent(client=client, language_code="zh") + agent = WorkflowDistillAgent(client=client, language_code="zh") captured_messages = [] @@ -398,7 +404,7 @@ def fake_call_llm(messages): return MOCK_CONTEXT_RESPONSE with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - agent.run(SAMPLE_EXPERIENCE_CONTEXT) + agent.run(SAMPLE_WORKFLOW_CONTEXT) system_content = captured_messages[0]["content"] assert "Simplified Chinese" in system_content @@ -406,7 +412,7 @@ def fake_call_llm(messages): def test_language_code_en_no_extra_instruction(self): client = self._mock_client() - agent = ExperienceDistillAgent(client=client, language_code="en") + agent = WorkflowDistillAgent(client=client, language_code="en") captured_messages = [] @@ -415,27 +421,45 @@ def fake_call_llm(messages): return MOCK_CONTEXT_RESPONSE with patch.object(agent, "_call_llm", side_effect=fake_call_llm): - agent.run(SAMPLE_EXPERIENCE_CONTEXT) + agent.run(SAMPLE_WORKFLOW_CONTEXT) system_content = captured_messages[0]["content"] assert "in English" in system_content assert "[LANGUAGE INSTRUCTION]" not in system_content -# ── _experience_filename ────────────────────────────────────────────────── +# ── _workflow_filename ────────────────────────────────────────────────── -class TestExperienceFilename: - def test_derives_from_workspace_name(self): - from data_formulator.routes.knowledge import _experience_filename - name = _experience_filename("Sales Analysis Pattern") +class TestWorkflowFilename: + def test_derives_from_title(self): + from data_formulator.routes.knowledge import _workflow_filename + name = _workflow_filename("Sales Analysis Pattern") assert name.endswith(".md") assert "sales-analysis-pattern" in name.lower() - def test_fallback_when_workspace_name_blank(self): - from data_formulator.routes.knowledge import _experience_filename - name = _experience_filename(" ") - assert name == "session-experience.md" + def test_fallback_when_title_blank(self): + from data_formulator.routes.knowledge import _workflow_filename + name = _workflow_filename(" ") + assert name == "session-workflow.md" + + def test_rejects_path_traversal(self): + from data_formulator.routes.knowledge import _workflow_filename + # An LLM-supplied name must never escape the workflows directory. + for evil in ("../../etc/passwd", "..\\..\\win", "/etc/shadow", "a/b/c"): + name = _workflow_filename(evil) + assert "/" not in name + assert "\\" not in name + assert ".." not in name + assert name.endswith(".md") + + def test_strips_reserved_and_control_chars(self): + from data_formulator.routes.knowledge import _workflow_filename + name = _workflow_filename('sales:report*?"<>|\x00 v1') + assert name.endswith(".md") + for ch in ':*?"<>|\x00': + assert ch not in name + assert name == "sales-report-v1.md" # ── API endpoint ────────────────────────────────────────────────────────── @@ -453,7 +477,7 @@ def app(self, tmp_path): _app.register_blueprint(knowledge_bp) register_error_handlers(_app) - (tmp_path / "knowledge" / "experiences").mkdir(parents=True) + (tmp_path / "knowledge" / "workflows").mkdir(parents=True) with patch("data_formulator.routes.knowledge.get_identity_id", return_value="test-user"), \ patch("data_formulator.routes.knowledge.get_user_home", return_value=tmp_path): @@ -464,14 +488,14 @@ def client(self, app): return app.test_client() def test_missing_context_returns_error(self, client): - resp = client.post("/api/knowledge/distill-experience", + resp = client.post("/api/knowledge/distill-workflow", json={"model": {"endpoint": "openai", "model": "gpt-4o"}}) data = resp.get_json() assert data["status"] == "error" def test_missing_model_returns_error(self, client): - resp = client.post("/api/knowledge/distill-experience", - json={"experience_context": SAMPLE_EXPERIENCE_CONTEXT}) + resp = client.post("/api/knowledge/distill-workflow", + json={"workflow_context": SAMPLE_WORKFLOW_CONTEXT}) data = resp.get_json() assert data["status"] == "error" @@ -483,9 +507,9 @@ def test_missing_events_returns_error(self, client): "workspace_name": "Demo", "threads": [], } - resp = client.post("/api/knowledge/distill-experience", + resp = client.post("/api/knowledge/distill-workflow", json={ - "experience_context": bad_context, + "workflow_context": bad_context, "model": {"endpoint": "openai", "model": "gpt-4o", "api_key": "test"}, }) data = resp.get_json() @@ -497,9 +521,9 @@ def test_missing_events_field_returns_error(self, client): "workspace_id": "ws-1", "workspace_name": "Demo", } # no 'threads' key - resp = client.post("/api/knowledge/distill-experience", + resp = client.post("/api/knowledge/distill-workflow", json={ - "experience_context": bad_context, + "workflow_context": bad_context, "model": {"endpoint": "openai", "model": "gpt-4o", "api_key": "test"}, }) data = resp.get_json() @@ -508,22 +532,22 @@ def test_missing_events_field_returns_error(self, client): def test_successful_distill(self, client, tmp_path): with patch("data_formulator.routes.agents.get_client") as mock_gc, \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ - patch("data_formulator.agents.agent_experience_distill.ExperienceDistillAgent.run", + patch("data_formulator.agents.agent_workflow_distill.WorkflowDistillAgent.run", return_value=MOCK_CONTEXT_RESPONSE): mock_gc.return_value = MagicMock() - resp = client.post("/api/knowledge/distill-experience", + resp = client.post("/api/knowledge/distill-workflow", json={ - "experience_context": SAMPLE_EXPERIENCE_CONTEXT, + "workflow_context": SAMPLE_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "model": "gpt-4o", "api_key": "test"}, }) data = resp.get_json() assert data["status"] == "success" - assert data["data"]["category"] == "experiences" + assert data["data"]["category"] == "workflows" assert data["data"]["path"].endswith(".md") # Verify file was written - exp_dir = tmp_path / "knowledge" / "experiences" + exp_dir = tmp_path / "knowledge" / "workflows" md_files = list(exp_dir.rglob("*.md")) assert len(md_files) >= 1 assert not (tmp_path / "agent-logs").exists() @@ -531,13 +555,13 @@ def test_successful_distill(self, client, tmp_path): def test_category_hint_creates_subdir(self, client, tmp_path): with patch("data_formulator.routes.agents.get_client") as mock_gc, \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ - patch("data_formulator.agents.agent_experience_distill.ExperienceDistillAgent.run", + patch("data_formulator.agents.agent_workflow_distill.WorkflowDistillAgent.run", return_value=MOCK_CONTEXT_RESPONSE): mock_gc.return_value = MagicMock() - resp = client.post("/api/knowledge/distill-experience", + resp = client.post("/api/knowledge/distill-workflow", json={ - "experience_context": SAMPLE_EXPERIENCE_CONTEXT, + "workflow_context": SAMPLE_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "model": "gpt-4o", "api_key": "test"}, "category_hint": "sales", }) diff --git a/tests/backend/knowledge/test_knowledge_store.py b/tests/backend/knowledge/test_knowledge_store.py index 2444195b..f69ce37c 100644 --- a/tests/backend/knowledge/test_knowledge_store.py +++ b/tests/backend/knowledge/test_knowledge_store.py @@ -5,11 +5,11 @@ Covers: - list_all, read, write, delete for each category -- path depth constraints (rules=flat, experiences=1 sub-dir) +- path depth constraints (rules=flat, workflows=1 sub-dir) - .md extension enforcement - ConfinedDir traversal rejection - front matter parsing and graceful degradation -- search: title, tags, filename, body matching + ranking + limit +- search: title, filename, body matching + ranking + limit - search skips alwaysApply rules (they are injected via system prompt) - tokenization: English stopwords, CJK/ASCII mixed splitting - scoring: partial token match, source discount, table_names boost @@ -73,21 +73,20 @@ def test_lists_rules(self, store, tmp_path): items = store.list_all("rules") assert len(items) == 1 assert items[0]["title"] == "ROI Calculation" - assert items[0]["tags"] == ["finance", "computation"] assert items[0]["path"] == "roi.md" assert items[0]["source"] == "manual" - def test_lists_experiences_in_subdirs(self, store, tmp_path): - exp_dir = tmp_path / "knowledge" / "experiences" / "cleaning" + def test_lists_workflows_in_subdirs(self, store, tmp_path): + exp_dir = tmp_path / "knowledge" / "workflows" / "cleaning" exp_dir.mkdir(parents=True) (exp_dir / "missing.md").write_text(SAMPLE_MD_SKILL, encoding="utf-8") - items = store.list_all("experiences") + items = store.list_all("workflows") assert len(items) == 1 assert items[0]["path"] == "cleaning/missing.md" def test_empty_category_returns_empty(self, store): - items = store.list_all("experiences") + items = store.list_all("workflows") assert items == [] def test_front_matter_title_fallback_to_stem(self, store, tmp_path): @@ -139,9 +138,9 @@ def test_preserves_existing_front_matter(self, store): content = store.read("rules", "fm.md") assert "title: ROI Calculation" in content - def test_writes_experiences_in_subdir(self, store, tmp_path): - store.write("experiences", "cleaning/handle-missing.md", SAMPLE_MD_SKILL) - assert (tmp_path / "knowledge" / "experiences" / "cleaning" / "handle-missing.md").exists() + def test_writes_workflows_in_subdir(self, store, tmp_path): + store.write("workflows", "cleaning/handle-missing.md", SAMPLE_MD_SKILL) + assert (tmp_path / "knowledge" / "workflows" / "cleaning" / "handle-missing.md").exists() # ── CRUD: delete ────────────────────────────────────────────────────────── @@ -169,12 +168,12 @@ def test_rules_subdir_rejected(self): with pytest.raises(ValueError, match="sub-directories"): KnowledgeStore.validate_path("rules", "sub/file.md") - def test_experiences_one_subdir_ok(self): - KnowledgeStore.validate_path("experiences", "cat/file.md") + def test_workflows_one_subdir_ok(self): + KnowledgeStore.validate_path("workflows", "cat/file.md") - def test_experiences_two_subdirs_rejected(self): + def test_workflows_two_subdirs_rejected(self): with pytest.raises(ValueError, match="one level"): - KnowledgeStore.validate_path("experiences", "cat/sub/file.md") + KnowledgeStore.validate_path("workflows", "cat/sub/file.md") def test_skills_rejected_as_invalid(self): with pytest.raises(ValueError, match="Invalid category"): @@ -228,7 +227,7 @@ def _setup_knowledge(self, store, tmp_path): rules_dir = tmp_path / "knowledge" / "rules" (rules_dir / "roi.md").write_text(SAMPLE_MD, encoding="utf-8") - exp_dir = tmp_path / "knowledge" / "experiences" / "cleaning" + exp_dir = tmp_path / "knowledge" / "workflows" / "cleaning" exp_dir.mkdir(parents=True) (exp_dir / "missing.md").write_text(SAMPLE_MD_SKILL, encoding="utf-8") @@ -237,11 +236,6 @@ def test_search_by_title(self, store): assert len(results) >= 1 assert results[0]["title"] == "Handle Missing Values" - def test_search_by_tags(self, store): - results = store.search("pandas") - assert len(results) >= 1 - assert results[0]["title"] == "Handle Missing Values" - def test_search_by_filename(self, store): results = store.search("missing") assert len(results) >= 1 @@ -269,7 +263,7 @@ def test_max_results_limit(self, store, tmp_path): assert len(results) <= 5 def test_search_filters_by_category(self, store): - results = store.search("ROI", categories=["experiences"]) + results = store.search("ROI", categories=["workflows"]) assert len(results) == 0 def test_search_skips_always_apply_rules(self, store, tmp_path): @@ -304,13 +298,12 @@ def test_partial_token_match_finds_results(self, store): assert results[0]["title"] == "Handle Missing Values" def test_table_names_boost(self, store, tmp_path): - """Entries tagged with a session table name get boosted.""" - exp_dir = tmp_path / "knowledge" / "experiences" / "analysis" + """Entries mentioning a session table name (title/body) get boosted.""" + exp_dir = tmp_path / "knowledge" / "workflows" / "analysis" exp_dir.mkdir(parents=True) (exp_dir / "sales-tip.md").write_text( - "---\ntitle: Sales Analysis Tips\n" - "tags: [sales_data, revenue]\nsource: manual\n---\n" - "When analysing sales, check for seasonality.\n", + "---\ntitle: Sales Analysis Tips\nsource: manual\n---\n" + "When analysing sales_data, check for seasonality.\n", encoding="utf-8", ) results = store.search("analysis tips", table_names=["sales_data"]) @@ -319,7 +312,7 @@ def test_table_names_boost(self, store, tmp_path): def test_non_manual_source_discounted(self, store, tmp_path): """Non-manual entries score lower than equivalent manual entries.""" - exp_dir = tmp_path / "knowledge" / "experiences" + exp_dir = tmp_path / "knowledge" / "workflows" (exp_dir / "auto-tip.md").write_text( "---\ntitle: Tip One\ntags: [tip]\nsource: distill\n---\nSome tip.\n", encoding="utf-8", @@ -328,7 +321,7 @@ def test_non_manual_source_discounted(self, store, tmp_path): "---\ntitle: Tip One\ntags: [tip]\nsource: manual\n---\nSome tip.\n", encoding="utf-8", ) - results = store.search("Tip One", categories=["experiences"]) + results = store.search("Tip One", categories=["workflows"]) assert len(results) == 2 assert results[0]["source"] == "manual" assert results[1]["source"] == "distill" @@ -472,7 +465,7 @@ def test_all_stopwords_returns_empty(self): class TestMatchScore: def test_single_token_title_hit(self): score = KnowledgeStore._match_score( - "ROI", "ROI Calculation", [], "roi", "", + "ROI", "ROI Calculation", "roi", "", ) assert score > 0 @@ -481,49 +474,49 @@ def test_partial_tokens_accumulate(self): score = KnowledgeStore._match_score( "quarterly sales trend", "Sales Trend Analysis", - [], "analysis", "", + "analysis", "", ) assert score > 0 def test_whole_string_bonus(self): full = KnowledgeStore._match_score( - "ROI", "ROI Calculation", [], "roi", "", + "ROI", "ROI Calculation", "roi", "", ) no_title = KnowledgeStore._match_score( - "ROI", "Something Else", [], "roi", "", + "ROI", "Something Else", "roi", "", ) assert full > no_title def test_source_discount(self): manual = KnowledgeStore._match_score( - "ROI", "ROI Guide", ["finance"], "roi", "", + "ROI", "ROI Guide", "roi", "", source="manual", ) auto = KnowledgeStore._match_score( - "ROI", "ROI Guide", ["finance"], "roi", "", + "ROI", "ROI Guide", "roi", "", source="distill", ) assert auto == pytest.approx(manual * 0.9) def test_table_names_boost(self): without = KnowledgeStore._match_score( - "analysis", "Analysis Tips", ["sales_data"], "tips", "", + "analysis", "Analysis Tips", "tips", "about sales_data", ) with_tn = KnowledgeStore._match_score( - "analysis", "Analysis Tips", ["sales_data"], "tips", "", + "analysis", "Analysis Tips", "tips", "about sales_data", table_names=["sales_data"], ) assert with_tn > without def test_no_match_returns_zero(self): score = KnowledgeStore._match_score( - "xyznonexistent", "ROI Calculation", ["finance"], "roi", "body text", + "xyznonexistent", "ROI Calculation", "roi", "body text", ) assert score == 0 def test_cjk_mixed_query_matches(self): """Chinese+English query should match via extracted ASCII tokens.""" score = KnowledgeStore._match_score( - "帮我分析ROI", "ROI Calculation", ["finance"], "roi", "", + "帮我分析ROI", "ROI Calculation", "roi", "", ) assert score > 0 diff --git a/tests/backend/routes/test_knowledge_routes.py b/tests/backend/routes/test_knowledge_routes.py index ddc2b7ab..f5ac69ff 100644 --- a/tests/backend/routes/test_knowledge_routes.py +++ b/tests/backend/routes/test_knowledge_routes.py @@ -167,7 +167,7 @@ def test_delete_nonexistent(self, client): class TestKnowledgeSearch: def test_search_returns_results(self, client, tmp_path): - exp_dir = tmp_path / "knowledge" / "experiences" / "finance" + exp_dir = tmp_path / "knowledge" / "workflows" / "finance" exp_dir.mkdir(parents=True, exist_ok=True) (exp_dir / "roi.md").write_text(SAMPLE_MD, encoding="utf-8") @@ -191,7 +191,7 @@ def test_search_invalid_category(self, client): assert data["status"] == "error" def test_search_filters_by_category(self, client, tmp_path): - exp_dir = tmp_path / "knowledge" / "experiences" / "finance" + exp_dir = tmp_path / "knowledge" / "workflows" / "finance" exp_dir.mkdir(parents=True, exist_ok=True) (exp_dir / "roi.md").write_text(SAMPLE_MD, encoding="utf-8") @@ -202,7 +202,7 @@ def test_search_filters_by_category(self, client, tmp_path): assert len(data["data"]["results"]) == 0 -SESSION_EXPERIENCE_CONTEXT = { +SESSION_WORKFLOW_CONTEXT = { "context_id": "ws-1", "workspace_id": "ws-1", "workspace_name": "Gasoline prices 2024", @@ -233,6 +233,7 @@ def test_search_filters_by_category(self, client, tmp_path): DISTILLED_MD = """\ --- subtitle: monthly sales aggregation +filename: monthly sales tags: [sales, time-series] created: 2026-05-06 updated: 2026-05-06 @@ -251,37 +252,37 @@ def test_search_filters_by_category(self, client, tmp_path): """ -class TestDistillExperience: - def test_distill_experience_from_context(self, client, tmp_path): +class TestDistillWorkflow: + def test_distill_workflow_from_context(self, client, tmp_path): with patch("data_formulator.routes.agents.get_client", return_value=object()), \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ patch( - "data_formulator.agents.agent_experience_distill." - "ExperienceDistillAgent.run", + "data_formulator.agents.agent_workflow_distill." + "WorkflowDistillAgent.run", return_value=DISTILLED_MD, ) as run: - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": SESSION_EXPERIENCE_CONTEXT, + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "success" - assert data["data"]["category"] == "experiences" - assert (tmp_path / "knowledge" / "experiences" / data["data"]["path"]).exists() + assert data["data"]["category"] == "workflows" + assert (tmp_path / "knowledge" / "workflows" / data["data"]["path"]).exists() assert not (tmp_path / "agent-logs").exists() run.assert_called_once() - def test_distill_experience_llm_timeout_returns_structured_error(self, client): + def test_distill_workflow_llm_timeout_returns_structured_error(self, client): with patch("data_formulator.routes.agents.get_client", return_value=object()), \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ patch( - "data_formulator.agents.agent_experience_distill." - "ExperienceDistillAgent.run", + "data_formulator.agents.agent_workflow_distill." + "WorkflowDistillAgent.run", side_effect=TimeoutError("request timed out"), ): - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": SESSION_EXPERIENCE_CONTEXT, + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) @@ -291,55 +292,60 @@ def test_distill_experience_llm_timeout_returns_structured_error(self, client): assert data["error"]["code"] == "LLM_TIMEOUT" assert data["error"]["retry"] is True - def test_distill_experience_missing_context(self, client): - resp = client.post("/api/knowledge/distill-experience", json={ + def test_distill_workflow_missing_context(self, client): + resp = client.post("/api/knowledge/distill-workflow", json={ "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "error" - def test_distill_experience_missing_threads(self, client): - bad_context = {k: v for k, v in SESSION_EXPERIENCE_CONTEXT.items() if k != "threads"} - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": bad_context, + def test_distill_workflow_missing_threads(self, client): + bad_context = {k: v for k, v in SESSION_WORKFLOW_CONTEXT.items() if k != "threads"} + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": bad_context, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "error" - def test_distill_experience_missing_workspace(self, client): - bad_context = {k: v for k, v in SESSION_EXPERIENCE_CONTEXT.items() + def test_distill_workflow_missing_workspace(self, client): + bad_context = {k: v for k, v in SESSION_WORKFLOW_CONTEXT.items() if k not in ("workspace_id", "workspace_name")} - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": bad_context, + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": bad_context, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "error" - def test_distill_session_overrides_title_with_workspace_name(self, client, tmp_path): - """Session-scoped distillation composes 'Experience from : '.""" + def test_distill_session_uses_descriptive_title(self, client, tmp_path): + """Session-scoped distillation uses the agent subtitle as the title.""" with patch("data_formulator.routes.agents.get_client", return_value=object()), \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ patch( - "data_formulator.agents.agent_experience_distill." - "ExperienceDistillAgent.run", + "data_formulator.agents.agent_workflow_distill." + "WorkflowDistillAgent.run", return_value=DISTILLED_MD, ): - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": SESSION_EXPERIENCE_CONTEXT, + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "success" path = data["data"]["path"] - # Filename is derived from the workspace name, not the LLM subtitle. - assert path == "gasoline-prices-2024.md" - saved = (tmp_path / "knowledge" / "experiences" / path).read_text(encoding="utf-8") - assert "title: 'Experience from Gasoline prices 2024: monthly sales aggregation'" in saved \ - or "title: \"Experience from Gasoline prices 2024: monthly sales aggregation\"" in saved \ - or "title: Experience from Gasoline prices 2024: monthly sales aggregation" in saved + # Filename is derived from the short agent-emitted `filename` hint, + # not the long descriptive title. + assert path == "monthly-sales.md" + saved = (tmp_path / "knowledge" / "workflows" / path).read_text(encoding="utf-8") + assert "title: monthly sales aggregation" in saved \ + or "title: 'monthly sales aggregation'" in saved \ + or "title: \"monthly sales aggregation\"" in saved + # No legacy "Workflow from :" prefix on the title. + assert "Workflow from" not in saved + # The filename hint is consumed, not persisted in the front matter. + assert "filename:" not in saved # Workspace stamps are present so the file can be looked up later. assert "source_workspace_id: ws-1" in saved assert "source_workspace_name: Gasoline prices 2024" in saved @@ -347,42 +353,46 @@ def test_distill_session_overrides_title_with_workspace_name(self, client, tmp_p assert "## Method" in saved def test_distill_session_upserts_existing_workspace_file(self, client, tmp_path): - """Re-distilling the same workspace overwrites the same file.""" + """Re-distilling the same workspace replaces the prior file.""" + second_md = DISTILLED_MD.replace( + "filename: monthly sales", + "filename: annual revenue", + ) with patch("data_formulator.routes.agents.get_client", return_value=object()), \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ patch( - "data_formulator.agents.agent_experience_distill." - "ExperienceDistillAgent.run", - return_value=DISTILLED_MD, + "data_formulator.agents.agent_workflow_distill." + "WorkflowDistillAgent.run", + side_effect=[DISTILLED_MD, second_md], ): - client.post("/api/knowledge/distill-experience", json={ - "experience_context": SESSION_EXPERIENCE_CONTEXT, + client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) - # Re-distill: workspace renamed, so the slug changes — old file - # should be removed in favour of the new one. - renamed = {**SESSION_EXPERIENCE_CONTEXT, "workspace_name": "Diesel 2024"} - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": renamed, + # Re-distill: the filename hint changes, so the slug changes — old + # file should be removed in favour of the new one (matched by + # source_workspace_id). + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "success" new_path = data["data"]["path"] - exp_dir = tmp_path / "knowledge" / "experiences" + exp_dir = tmp_path / "knowledge" / "workflows" # Stale slug deleted, new slug present. - assert not (exp_dir / "gasoline-prices-2024.md").exists() + assert not (exp_dir / "monthly-sales.md").exists() assert (exp_dir / new_path).exists() - assert new_path == "diesel-2024.md" + assert new_path == "annual-revenue.md" - def test_distill_session_skips_subtitle_double_prefix(self, client, tmp_path): - """Update-mode runs that re-emit a prefixed title don't double-prefix.""" - # Simulate a prior run where the LLM echoed an Experience-prefixed title + def test_distill_session_strips_legacy_title_prefix(self, client, tmp_path): + """Update-mode runs strip any legacy 'Workflow from :' prefix.""" + # Simulate a prior run where the LLM echoed a Workflow-prefixed title # without a subtitle. prior_md = ( "---\n" - "title: 'Experience from Gasoline prices 2024: prior insight'\n" + "title: 'Workflow from Gasoline prices 2024: prior insight'\n" "tags: [a]\n" "created: 2026-05-06\n" "updated: 2026-05-06\n" @@ -392,17 +402,18 @@ def test_distill_session_skips_subtitle_double_prefix(self, client, tmp_path): with patch("data_formulator.routes.agents.get_client", return_value=object()), \ patch("data_formulator.routes.agents.get_language_instruction", return_value=""), \ patch( - "data_formulator.agents.agent_experience_distill." - "ExperienceDistillAgent.run", + "data_formulator.agents.agent_workflow_distill." + "WorkflowDistillAgent.run", return_value=prior_md, ): - resp = client.post("/api/knowledge/distill-experience", json={ - "experience_context": SESSION_EXPERIENCE_CONTEXT, + resp = client.post("/api/knowledge/distill-workflow", json={ + "workflow_context": SESSION_WORKFLOW_CONTEXT, "model": {"endpoint": "openai", "key": "x", "model": "gpt"}, }) data = resp.get_json() assert data["status"] == "success" - saved = (tmp_path / "knowledge" / "experiences" / data["data"]["path"]).read_text(encoding="utf-8") - # The "Experience from ..." prefix is stripped before re-prefixing. - assert saved.count("Experience from") == 1 + saved = (tmp_path / "knowledge" / "workflows" / data["data"]["path"]).read_text(encoding="utf-8") + # The legacy "Workflow from ..." prefix is fully stripped. + assert "Workflow from" not in saved + assert "prior insight" in saved From 1571113ac7b74ce1a0c32b434a9576d6fdce2b40 Mon Sep 17 00:00:00 2001 From: y-agent-ai Date: Sat, 30 May 2026 18:22:30 +0800 Subject: [PATCH 07/29] refactor(loading): Refactor AnvilLoader and add custom parameter support 1. Add custom property support for height , label , and sx to AnvilLoader 2. Replace globally hardcoded loading text with customizable label parameter 3. Optimize loading overlay styles with new frosted glass background effect 4. Unify loading state display in App.tsx and VisualizationView --- src/app/App.tsx | 2 +- src/components/AnvilLoader.tsx | 46 +++++++++++++++++++++------------ src/views/VisualizationView.tsx | 10 ++++--- 3 files changed, 36 insertions(+), 22 deletions(-) diff --git a/src/app/App.tsx b/src/app/App.tsx index 17898f0c..22d76510 100644 --- a/src/app/App.tsx +++ b/src/app/App.tsx @@ -1253,7 +1253,7 @@ export const AppFC: FC = function AppFC(appProps) { {configLoaded && authChecked ? ( ) : ( - + )} {migrationBrowserId && ( ; +} + +export function AnvilLoader({ height = '100vh', label, sx }: AnvilLoaderProps) { return ( - - loading data formulator... - + {label !== undefined && ( + + {label} + + )} ); } diff --git a/src/views/VisualizationView.tsx b/src/views/VisualizationView.tsx index 7b6d18b4..585eba79 100644 --- a/src/views/VisualizationView.tsx +++ b/src/views/VisualizationView.tsx @@ -15,7 +15,6 @@ import { ListItemIcon, ListItemText, MenuItem, - LinearProgress, Card, ListSubheader, Menu, @@ -37,6 +36,7 @@ import _ from 'lodash'; import { borderColor, transition } from '../app/tokens'; import { WritingIndicator } from '../components/FunComponents'; +import { AnvilLoader } from '../components/AnvilLoader'; import ButtonGroup from '@mui/material/ButtonGroup'; @@ -1099,10 +1099,12 @@ export const ChartEditorFC: FC<{}> = function ChartEditorFC({}) { return {synthesisRunning ? - + : ''} {chartUnavailable ? "" : chartResizer} {content} From b9abafb163d59c8c4165075d8f573345b4d70235 Mon Sep 17 00:00:00 2001 From: cat0825 <1759138827@qq.com> Date: Sun, 31 May 2026 14:02:02 +0800 Subject: [PATCH 08/29] test: keep zh locale keys aligned --- src/i18n/locales/zh/dataLoading.json | 1 + tests/frontend/unit/app/i18nLocales.test.ts | 30 +++++++++++++++++++++ 2 files changed, 31 insertions(+) create mode 100644 tests/frontend/unit/app/i18nLocales.test.ts diff --git a/src/i18n/locales/zh/dataLoading.json b/src/i18n/locales/zh/dataLoading.json index 4eab6fcf..6ffe595d 100644 --- a/src/i18n/locales/zh/dataLoading.json +++ b/src/i18n/locales/zh/dataLoading.json @@ -39,6 +39,7 @@ "rowLimit": "行数限制", "loadSelected": "加载选中的表", "loadedCount": "✓ 已加载 {{count}} 张表", + "loadedCount_plural": "✓ 已加载 {{count}} 张表", "preview": "预览", "hidePreview": "收起", "previewing": "正在预览...", diff --git a/tests/frontend/unit/app/i18nLocales.test.ts b/tests/frontend/unit/app/i18nLocales.test.ts new file mode 100644 index 00000000..dd6c9933 --- /dev/null +++ b/tests/frontend/unit/app/i18nLocales.test.ts @@ -0,0 +1,30 @@ +import { describe, expect, it } from "vitest"; + +import en from "../../../../src/i18n/locales/en"; +import zh from "../../../../src/i18n/locales/zh"; + +type TranslationValue = string | Record; +type TranslationMap = Record; + +function collectKeys(value: TranslationMap, prefix = ""): Set { + const keys = new Set(); + + for (const [key, child] of Object.entries(value)) { + const nextPrefix = prefix ? `${prefix}.${key}` : key; + if (typeof child === "string") { + keys.add(nextPrefix); + } else { + for (const childKey of collectKeys(child, nextPrefix)) { + keys.add(childKey); + } + } + } + + return keys; +} + +describe("i18n locale bundles", () => { + it("keeps Simplified Chinese translation keys aligned with English", () => { + expect(collectKeys(zh)).toEqual(collectKeys(en)); + }); +}); From 748a30ce45f8e01b95388aced0b2e27106d70b69 Mon Sep 17 00:00:00 2001 From: Chenglong Wang Date: Sun, 31 May 2026 12:26:36 -0700 Subject: [PATCH 09/29] bug fix and clean up --- .../agents/agent_workflow_distill.py | 190 ++++++++++-------- .../data_loader/sample_datasets_loader.py | 13 +- src/app/dfSlice.tsx | 10 + src/components/LoadPlanCard.tsx | 51 ++++- src/i18n/locales/en/common.json | 4 +- src/i18n/locales/en/dataLoading.json | 2 + src/i18n/locales/zh/dataLoading.json | 2 + src/views/DataLoadingChat.tsx | 24 ++- src/views/DataSourceSidebar.tsx | 28 ++- src/views/EncodingShelfCard.tsx | 60 ++++-- src/views/KnowledgePanel.tsx | 6 +- src/views/VisualizationView.tsx | 15 +- 12 files changed, 263 insertions(+), 142 deletions(-) diff --git a/py-src/data_formulator/agents/agent_workflow_distill.py b/py-src/data_formulator/agents/agent_workflow_distill.py index 3f3d9c6d..0d86aa78 100644 --- a/py-src/data_formulator/agents/agent_workflow_distill.py +++ b/py-src/data_formulator/agents/agent_workflow_distill.py @@ -30,9 +30,28 @@ SYSTEM_PROMPT = """\ You are a workflow distiller. Given the chronological events of a data -analysis session plus an optional user instruction, extract a short, -**replayable workflow** that captures *what the user wanted and got* — so -the same analysis can be reproduced later on a similarly-shaped dataset. +analysis session plus an optional user distillation instruction, extract a **replayable +workflow** that captures *what the user wanted and got* — and write it at +TWO levels so it can be reused in two different situations: + +1. An **Abstract workflow** — dataset-independent. The underlying analytical + pattern, stripped of this dataset's subject matter: the sequence of + questions, computations, and chart kinds, phrased in domain-neutral terms. + Following it on a *different and possibly very differently-shaped* dataset + should walk the same process and arrive at structurally similar + visualizations. +2. A **Concrete workflow** — for *similar* data (same shape, only minor + differences — a different period, region, or filter). It names the real + fields, aggregations, filters, and chart encodings used here, so the + analysis can be replayed closely with minimal thought. + +Both describe the SAME analysis at different distances. They should be +consistent, but they do NOT need an exact 1:1 step mapping — let each be as +long as it needs (typically 3-7 steps each). + +Where the analysis hinges on a few choices a user might change on replay (a +period, a filter, a top-N), surface them as named **parameters** with +`{{token}}` placeholders in the steps — see the `## Parameters` section below. The session contains one or more threads (separate analysis branches in the same session) each rendered under a `### Thread N` header. When @@ -47,38 +66,20 @@ (followed by columns, row count, sample, and code). - `create_chart` — a chart emitted on a table (mark + encoding summary). -Your job is to recover the **ordered list of requests** the user actually -wanted, and the outputs (tables/charts) they ended up keeping. Beyond the -concrete steps, also distill the analysis at TWO levels of abstraction so -it can be reused later: -- **Adapting to similar data** (concrete) — how to rerun essentially the - same analysis on a near-identical dataset, e.g. the business report for - a different month, region, or product line. Same shape and intent, only - the specific inputs/filters change. -- **Generalizing to other data** (abstract, dataset-agnostic) — the - underlying analytical pattern, independent of this domain: the kinds of - questions, computations, and charts involved, phrased so they transfer - to a different domain or a differently-shaped dataset. - CRITICAL extraction rules — keep only what the user wanted and got: -- Each step = one user request, written in plain language. Say BOTH the - question being explored AND what was produced to answer it — including - the chart that was created and the key fields it uses (e.g. "Ask how - sales trend over time, and plot monthly total sales as a line chart"; - "Compare regions by breaking revenue down per region as a sorted bar - chart"). Order them as the analysis progressed. +- Recover the ORDERED list of requests the user actually wanted, and the + outputs (tables/charts) they kept. Each step states BOTH the question + explored AND what was produced to answer it — including the chart and the + key fields it uses. - DROP corrective back-and-forth. If the user changed their mind ("no, it should be…", "actually use median instead"), keep ONLY the final resolved intent — not the wrong first attempt or the correction. - DROP abandoned work. If a chart or table was created and then deleted or never kept, leave it out entirely. - DROP mechanics. Do NOT include error-repair loops, dtype fixes, tool - call noise, or low-level code. Describe intent, not implementation. -- Do NOT lean on code or exact column names unless a name is essential to - the request's meaning. Keep steps dataset-agnostic where possible so - they replay on a new slice of similar data. -- Capture genuine gotchas separately as short notes (advisory warnings to - carry forward), NOT as steps to re-perform. + call noise, or low-level code dumps. Describe intent, not implementation. +- Capture genuine gotchas as short Notes (advisory warnings to carry + forward), NOT as steps to re-perform. If a user instruction is provided, let it steer what to keep or emphasise. @@ -86,8 +87,8 @@ ``` --- -subtitle: -filename: +subtitle: +filename: created: updated: source: distill @@ -96,74 +97,85 @@ ## Goal - -## Steps -1. -2. +what it produces. This is where the dataset-grounded explanation belongs — +you MAY name the real subject here (e.g. "originally distilled from a +monthly gasoline-price session").> + +## Parameters + +- `{{period}}` — the time range analysed; used here: 2024; on replay: ask. +- `{{top_n}}` — how many top categories to keep; used here: 10; on replay: keep. +- `{{region}}` — geographic filter applied; used here: National; on replay: ask. + +## Abstract workflow + +1. +2. 3. <…> -## Adapting to similar data - - -## Generalizing to other data - +## Concrete workflow + +1. +2. <…> ## Notes +analysis on new data — e.g. "sort by time before computing period-over-period +change". Omit this section entirely if there is nothing worth warning about.> ``` Rules: -- Subtitle must DESCRIBE what the workflow is about in PLAIN LANGUAGE that - a non-expert can fully understand at a glance, so they can decide - whether to replay it on new data. Favor clarity over brevity: it can be - a full sentence (up to ~25 words) if that makes the analysis genuinely - understandable. Write it like you would explain the analysis to a - colleague in one breath, covering the subject and the main thing you do - with it. The hosting application uses this subtitle directly as the - workflow's display title, so make it self-contained and do NOT prefix it - with the session name. - - Start with a concrete action verb (Plot, Compare, Break down, Rank, - Track, Summarize, Find…). - - Name the real-world subject in everyday words (sales, revenue, - customers, events), NOT the internal mechanics or derived-column - names you happened to create. - - AVOID abstract or technical jargon and invented noun-phrases - ("deltas", "composition", "window", "distribution shift"). If a - technique matters, phrase it plainly ("change from one period to the - next" instead of "deltas"). - Good: "Plot monthly sales over time and compare each year against the - previous one to spot volatile periods". - "Break revenue down by region and show how each region - contributes to the total as a stacked area chart". - "Track how many events happen in each time window and what kinds - of events make up each window". - Bad: "Time series analysis". "Data workflow". "Chart exploration". - "Event window deltas with composition". "Distribution shift inspection". +- The subtitle is the workflow's display TITLE. Make it ABSTRACT and + library-friendly: name the *kind of analysis* — a technique plus a GENERIC + subject (KPI, metric, category, event, cohort) — so someone browsing the + workflow library can tell whether this is the KIND of analysis they want to + reuse. Do NOT pin it to this dataset's specific subject, period, or column + names, and do NOT prefix it with the session name. + - Pair a real technique with a generic subject; avoid bare category words. + Good: "Year-over-year KPI volatility analysis". + "Category contribution-to-total breakdown". + "Time-windowed event composition analysis". + Bad: "Plot monthly gasoline prices in 2024 and compare each year". (too specific) + "Time series analysis". "Data workflow". "Chart exploration". (too vague) + The dataset-grounded, full-sentence explanation goes in `## Goal`, NOT the title. - Filename must be a SHORT (2-5 word) lowercase name for the file — just - the core subject and action, e.g. "monthly sales trend", "region revenue - breakdown". No dates, no file extension, no session name. It is only - used to name the file on disk; the descriptive subtitle is what users see. -- Steps must be ordered and reproducible. Each step should make clear the - question being explored and the chart/output produced to answer it. -- "Adapting to similar data" stays close to this analysis (same domain, - same shape) — only the concrete inputs change. "Generalizing to other - data" must be domain-neutral: strip out this dataset's subject matter and - describe only the transferable analytical pattern (question types, - computations, chart kinds). Do NOT just repeat the steps in either - section; add genuine reuse guidance. Keep each section brief. -- Be as long as the analysis needs — do not omit meaningful steps, - questions, or charts just to stay short. Stay focused, but completeness - matters more than brevity. + the technique/subject, e.g. "kpi volatility analysis", "region revenue + breakdown". No dates, no file extension, no session name. It only names the + file on disk; the subtitle is what users see. +- Abstract workflow must be domain-neutral — strip this dataset's subject + matter and column names; describe only the transferable pattern (question + types, computations, chart kinds). Concrete workflow must be runnable on a + near-identical dataset: real field names, the aggregation, the filter to + vary, the chart mark + key encodings. Do NOT have the two sections merely + repeat each other — each adds its own grain of reuse guidance. +- Parameters are optional and a judgment call: surface only the FEW knobs + that materially change the outcome and that a user would revisit on replay + (often 0-4). When in doubt, leave the value inline — a spurious `{{token}}` + is worse than none. Knobs may be run-specific (period, region, top-N — + usually `ask`) or dataset-specific (a domain value/column — usually `keep`, + and may be skipped in the Abstract workflow). Every `{{token}}` in the steps + must be listed in `## Parameters` and vice versa. +- Steps in both sections must be ordered and reproducible. +- Be as long as the analysis needs — do not omit meaningful steps, questions, + or charts just to stay short. Stay focused, but completeness matters more + than brevity. - No raw data, PII, secrets, or specific values unless essential to a request. - Write the subtitle, headings, and body in {output_language}. YAML front-matter keys stay in English. diff --git a/py-src/data_formulator/data_loader/sample_datasets_loader.py b/py-src/data_formulator/data_loader/sample_datasets_loader.py index 6c3267cf..a84b8302 100644 --- a/py-src/data_formulator/data_loader/sample_datasets_loader.py +++ b/py-src/data_formulator/data_loader/sample_datasets_loader.py @@ -25,7 +25,10 @@ import pyarrow as pa from data_formulator.data_loader.external_data_loader import ExternalDataLoader -from data_formulator.datalake.parquet_utils import df_to_safe_records +from data_formulator.datalake.parquet_utils import ( + df_to_safe_records, + sanitize_dataframe_for_arrow, +) logger = logging.getLogger(__name__) @@ -231,7 +234,13 @@ def fetch_data_as_arrow( logger.info("Returning %d / %d rows from sample dataset: %s", len(df), self._last_total_rows, source_table) - return pa.Table.from_pandas(df, preserve_index=False) + # Public sample JSON/CSV files frequently contain mixed-type object + # columns (e.g. movies.json's ``Title`` holds both strings and + # numeric values), which makes ``pa.Table.from_pandas`` raise + # ArrowTypeError. Coerce such columns to a consistent type first. + return pa.Table.from_pandas( + sanitize_dataframe_for_arrow(df), preserve_index=False + ) # ------------------------------------------------------------------ # Internal: cached full-dataset fetch diff --git a/src/app/dfSlice.tsx b/src/app/dfSlice.tsx index 89b075ab..2ceb55aa 100644 --- a/src/app/dfSlice.tsx +++ b/src/app/dfSlice.tsx @@ -245,6 +245,9 @@ export interface DataFormulatorState { /** Whether the data source sidebar is expanded (true) or collapsed to rail (false) */ dataSourceSidebarOpen: boolean; + /** Which data source sidebar tab is active. Persisted so it survives session refresh. */ + dataSourceSidebarTab: 'sources' | 'sessions' | 'knowledge'; + /** * One-shot signal asking the sidebar to focus a specific connector * (open the sidebar, switch to sources tab, expand + scroll-into-view @@ -322,6 +325,8 @@ const initialState: DataFormulatorState = { dataSourceSidebarOpen: false, + dataSourceSidebarTab: 'sources', + focusedConnectorId: undefined, } @@ -762,12 +767,16 @@ export const dataFormulatorSlice = createSlice({ viewMode: state.viewMode, dataLoaderConnectParams: state.dataLoaderConnectParams, dataSourceSidebarOpen: state.dataSourceSidebarOpen, + dataSourceSidebarTab: state.dataSourceSidebarTab, activeWorkspace: action.payload, }; }, setDataSourceSidebarOpen: (state, action: PayloadAction) => { state.dataSourceSidebarOpen = action.payload; }, + setDataSourceSidebarTab: (state, action: PayloadAction<'sources' | 'sessions' | 'knowledge'>) => { + state.dataSourceSidebarTab = action.payload; + }, /** * Ask the data-source sidebar to focus a specific connector. * Opens the sidebar (if collapsed) and stores the target id; the @@ -870,6 +879,7 @@ export const dataFormulatorSlice = createSlice({ activeWorkspace: saved.activeWorkspace ?? state.activeWorkspace ?? null, dataSourceSidebarOpen: state.dataSourceSidebarOpen, + dataSourceSidebarTab: state.dataSourceSidebarTab, // Reset display-rows tick so dependent components re-fetch. displayRowsTick: 0, diff --git a/src/components/LoadPlanCard.tsx b/src/components/LoadPlanCard.tsx index 9b60effd..b91e54ab 100644 --- a/src/components/LoadPlanCard.tsx +++ b/src/components/LoadPlanCard.tsx @@ -16,8 +16,14 @@ import type { LoadPlan, LoadPlanCandidate, PendingTableLoad } from './ComponentT interface LoadPlanCardProps { plan: LoadPlan; - onConfirm: (selected: LoadPlanCandidate[]) => void; + onConfirm: (selected: LoadPlanCandidate[], opts?: { newWorkspace?: boolean }) => void; confirmed?: boolean; + /** When true, a workspace with existing data is already open, so the + * destination of the load is ambiguous. We then offer two explicit + * actions: add to the current workspace, or load into a fresh one. + * When false (empty/new workspace), a single "Load selected" button + * loads directly with no ambiguity. */ + canLoadInNewWorkspace?: boolean; } // Plans this small auto-expand each row's preview on first render so the @@ -48,7 +54,7 @@ const formatFilterValue = (value: any) => { return Array.isArray(value) ? value.join(', ') : String(value); }; -export const LoadPlanCard: React.FC = ({ plan, onConfirm, confirmed }) => { +export const LoadPlanCard: React.FC = ({ plan, onConfirm, confirmed, canLoadInNewWorkspace }) => { const theme = useTheme(); const { t } = useTranslation(); const [selection, setSelection] = useState>( @@ -143,12 +149,12 @@ export const LoadPlanCard: React.FC = ({ plan, onConfirm, con fetchPreview(candidate, idx); }; - const handleConfirm = async () => { + const handleConfirm = async (newWorkspace = false) => { const selected = plan.candidates.filter((c, i) => selection[i] && !c.resolutionError); if (selected.length === 0) return; setLoading(true); try { - await onConfirm(selected); + await onConfirm(selected, { newWorkspace }); } finally { setLoading(false); } @@ -257,12 +263,47 @@ export const LoadPlanCard: React.FC = ({ plan, onConfirm, con defaultValue: '✓ Loaded', })} + ) : canLoadInNewWorkspace ? ( + // A workspace with data is already open — make the load + // destination explicit rather than silently appending. + <> + + + ) : (