AI-Suggested Single Mappings

[brandonopened]

AI-Suggested Mappings on the MDR Mapping Canvas

Is your feature request related to a problem? Please describe.
Mapping a new source data model to LIF (or any target) is the most time-consuming part of onboarding a data source. Today users have to manually scan
dozens to hundreds of source/target attributes and drag wires one-by-one, which is repetitive for the obvious cases (e.g. firstName → firstName) and
easy to miss when names differ slightly. It's the #1 friction point reported in workshops and the dominant blocker to onboarding new sources at demo
customers.

Describe the solution you'd like
Add a "Suggest mappings" action to the MDR mapping canvas that uses an LLM to propose attribute-to-attribute mappings between the loaded source and
target models. Suggestions render as dashed wires distinct from confirmed mappings. Clicking a suggested wire opens a small popover showing confidence,
source → target, a one-sentence reason, and approve (✓) / reject (✗) buttons. Approving creates a normal Transformation (identical in shape to a
hand-drawn one); rejecting removes the dashed wire. Suggestions are opt-in via an API key, never auto-commit, and the human always reviews each one.

Describe alternatives you've considered

• Fuzzy name matching only (no LLM): cheap, but misses semantic synonyms, type/unit differences, and won't propose transform expressions.
• Bulk import via CSV mapping spec: useful for power users but doesn't help with first-time discovery and still requires authoring the mapping somewhere.
• Auto-apply high-confidence matches: faster but unsafe — silent mappings are hard to audit and erode trust.
• Separate "AI mode" view: would split the workflow; inline suggestions on the existing canvas keeps everything in one place.

Additional context
A working prototype already exists on the ai-suggest branch (commits f4011bb and following). It uses Claude via the ANTHROPIC_API_KEY env var, renders https://github.com/brandonopened/lif-core/tree/ai-suggest
dashed purple wires for suggestions, and has the approve/reject popover wired up to the existing Transformations create endpoint. Open questions for the
team: should rejections persist across sessions, where the button should live in the toolbar, and whether on-prem/local-model support is needed for
customers who can't send schema metadata to a hosted LLM.

[brandonopened]

⏺ Endpoint is live at /suggest_mappings/. Here's a summary of what's working on the ai-suggest branch:

What's working on ai-suggest

Backend

• New endpoint POST /suggest_mappings/ (bases/lif/mdr_restapi/suggest_mappings_endpoint.py) — accepts a "selected" attribute and a list of candidates from
  the opposite side, returns ranked suggestions with confidence (0–1) and a 1-sentence reason.
• LLM integration — uses the official anthropic SDK with claude-sonnet-4-20250514. Reads ANTHROPIC_API_KEY from env; cleanly errors out if unset.
• Tests — test/bases/lif/mdr_restapi/test_suggest_mappings_endpoint.py (357 lines) covers the request/response contract and basic behaviors.

Frontend

• SuggestionPopover component — shows confidence %, src→tgt, reason, ✓/✗ buttons. Now auto-flips below the click and clamps horizontally so it can't be cut
  off (today's fix).
• useSuggestedMappings hook — manages suggestion state, dedupes stale requests (requestIdRef), filters out already-mapped attributes, caps candidates at 200.
• Wires renders dashed curves for suggestions, distinct from confirmed mappings.
  • MappingsView integration — clicking a source/target field triggers suggestion fetch; clicking a dashed wire opens the popover; ✓ creates a Transformation
    via the existing endpoint; ✗ removes the suggestion locally.

End-to-end flow that works today

1. User clicks a source or target attribute → loading state appears.
2. Backend calls Claude with selected field + candidates + any existing mappings as context.
3. Dashed purple wires render for returned suggestions, ranked by confidence.
4. User clicks a wire → popover with reason + approve/reject.
5. ✓ → creates a Transformation (default expression = ), wire becomes solid.
6. ✗ → wire disappears; no DB write.

Known follow-ups (not on branch yet)

• suggest_mappings_endpoint isn't include_router'd in core.py (must be auto-discovered or mounted elsewhere — works live but worth making explicit).
• No Source = "AI_SUGGESTED" flag on Transformations yet — approved AI mappings are indistinguishable from manual ones.
• Rejections are session-only; no persistence.
• No bulk-approve / confidence-threshold UI.