feat: replace faqqer with multi-channel AI help bot (AnythingLLM + Discord + Telegram) by 0xPepeSilvia · Pull Request #3 · tari-project/faqqer

0xPepeSilvia · 2026-04-14T19:59:25Z

Closes #1

Summary

This PR replaces the single-file faqqer_bot.py (direct OpenAI completions) with a production-ready multi-channel AI help bot backed by AnythingLLM's RAG pipeline. Telegram and Discord Q&A bots are deployed alongside the existing blockchain stats and customer-analysis scheduled jobs, all orchestrated via Docker Compose.

Acceptance criteria

#	Criterion	Status	File(s)
1	Framework selected and documented	✅	`docs/ARCHITECTURE.md`
2	Responds on Telegram grounded in Tari KB	✅	`bridge/telegram_bot.py` → `bridge/rag_client.py`
3	Responds on Discord grounded in Tari KB	✅	`bridge/discord_bot.py` → `bridge/rag_client.py`
4	KB updatable without code changes	✅	`bridge/init_kb.py`, volume mount in `docker-compose.yml`
5	Learns from feedback	✅	`bridge/feedback.py` + `promote_to_kb()` in `rag_client.py`
6	Preserves blockchain_job + customer_analysis_job	✅	`bridge/jobs.py` wraps both via `importlib` (no edits to originals)
7	Dockerized	✅	`docker-compose.yml` + `bridge/Dockerfile`

Framework rationale — why AnythingLLM

Admin UI: non-technical admins can manage documents, inspect conversations and rotate models at :3001 without touching code
Model-agnostic: switch from OpenAI to Anthropic/Ollama/Azure by changing one env var; no code changes
Stable REST API: /api/v1/workspace/{slug}/chat is versioned and thin — the whole RAG client is ~194 lines
Docker-native: official mintplexlabs/anythingllm image; LanceDB vector store is built-in, no extra infra
KB hot-reload: drop files in the volume-mounted faqs/ dir and restart bridge; manifest prevents duplicates

Architecture

                ┌──────────────────────────────────────────┐
                │           docker-compose stack            │
                │                                           │
 Telegram ──────►                              ┌──────────┐ │
                │   bridge worker  ◄─ REST ───► AnythingLLM│ │
 Discord  ──────►   (async tasks)              │ + LanceDB │ │
                │        │                     └──────────┘ │
                │   APScheduler                      ▲       │
                │   (blockchain + analysis jobs)     │       │
                │        │                    faqs/ volume   │
                │   Telethon                  (hot-reload)   │
                └──────────────────────────────────────────┘

Quick deploy

cp .env.example .env          # fill OPENAI_API_KEY, bot tokens
docker compose up -d anythingllm
# visit http://localhost:3001 → Settings → API Keys → generate key
# add key to .env as ANYTHINGLLM_API_KEY
docker compose up -d bridge
docker compose logs -f bridge  # confirm startup

Test results

Test 1 — syntax + import:

python -m py_compile bridge/config.py bridge/rag_client.py bridge/telegram_bot.py \
  bridge/discord_bot.py bridge/feedback.py bridge/jobs.py bridge/init_kb.py bridge/main.py
# → PASS (no output)

Test 2 — docker compose config:
Docker not available in the CI environment; compose YAML was manually validated: correct version, all service keys present, volumes/networks defined, env_file + environment blocks syntactically valid, no port conflicts.

Test 3 — adversarial review + fixes applied:

Scenario	Finding	Fix
`ANYTHINGLLM_API_KEY` missing	`config.py` raises `RuntimeError` with clear message; `main.py` catches it, logs and exits with code 2	Already correct
AnythingLLM unreachable	`rag_client.wait_until_ready()` polls with exponential back-off up to 5 min; `main.py` exits with code 3 if never reachable	Already correct
Discord message > 2000 chars	No truncation — raw `answer` passed to `.reply()` / `.followup.send()`	Fixed: added `_truncate()` helper in `discord_bot.py` applied at all send sites
Feedback on message bot didn't send	`feedback.get()` returns `None`; all approve handlers check for `None` and reply gracefully	Already correct; also added try/except around `_record_pair` in `on_message` path
Missing deps in `requirements.txt`	`SQLAlchemy` (required by `APScheduler.SQLAlchemyJobStore`) and `openai` (required by `customer_analysis_job`) were absent	Added `SQLAlchemy==2.0.36` and `openai==1.54.4`

🤖 Generated with Claude Code

Bounty payout

If this PR is accepted as a bounty submission, please direct funds to the following Tari wallet address:

123JbAxCZ4Vrh4jCjFLXTu4z8sLeggUR87fejwNunsB55NMBSH9RpkjNiw1WL2GkKr8JhqRZhbrsSREchGanchw2Nhb

gemini-code-assist

Code Review

This pull request upgrades the Tari AI Help Bot to version 2, implementing a Retrieval-Augmented Generation (RAG) pipeline powered by AnythingLLM. The new architecture features a multi-channel bridge worker supporting Telegram and Discord, automated knowledge base seeding from local files, and a feedback system for admin-approved learning. It also integrates legacy scheduled jobs using APScheduler. Review feedback identifies several improvement opportunities: optimizing performance by moving the HTTP client instantiation outside of retry loops, enhancing Discord bot reliability by utilizing member data directly from event payloads, and preventing potential API errors in the Telegram bot by implementing message length truncation.

gemini-code-assist · 2026-04-14T20:02:10Z

+        async for attempt in AsyncRetrying(
+            stop=stop_after_attempt(3),
+            wait=wait_exponential(min=1, max=8),
+            retry=retry_if_exception_type((httpx.HTTPError,)),
+            reraise=True,
+        ):
+            with attempt:
+                async with httpx.AsyncClient(
+                    timeout=self._timeout, headers=self._headers
+                ) as client:
+                    resp = await client.post(url, json=payload)
+                    if resp.status_code >= 500:
+                        raise httpx.HTTPError(f"upstream {resp.status_code}")
+                    if resp.status_code >= 400:
+                        raise RAGError(
+                            f"AnythingLLM rejected the request: "
+                            f"{resp.status_code} {resp.text[:200]}"
+                        )
+                    data = resp.json()


The httpx.AsyncClient is being instantiated inside the retry loop. This is inefficient as it creates a new client and connection pool for every retry attempt. It should be moved outside the AsyncRetrying block to properly leverage connection pooling across retries.

References

Reuse client sessions to benefit from connection pooling and improve performance.

Fixed in 99fab49: httpx.AsyncClient is now instantiated once before the AsyncRetrying loop and reused across retry attempts.

gemini-code-assist · 2026-04-14T20:02:10Z

+            guild = bot.get_guild(payload.guild_id) if payload.guild_id else None
+            member = guild.get_member(payload.user_id) if guild else None
+            if not self._is_admin(member):
+                return


In on_raw_reaction_add, the payload object already contains the member object for guild events. Using payload.member is more reliable than guild.get_member, especially since the members intent is not enabled, which would likely result in cache misses for users not already known to the bot.

References

Use available event payload data to avoid unnecessary API calls or cache lookups.

Fixed in 99fab49: on_raw_reaction_add now reads payload.member first (populated directly by the Discord gateway event) and only falls back to guild.get_member() when it's None.

gemini-code-assist · 2026-04-14T20:02:10Z

+        answer = await self._answer(question)
+        sent = await msg.reply_text(answer)


Telegram has a message length limit of 4096 characters. If the AI response exceeds this limit, the bot will fail to send the reply. Truncation should be applied to ensure the message is within Telegram's limits, similar to the implementation in the Discord bot.

References

Handle platform-specific message length limits to prevent API errors.

Fixed in 99fab49: added a _truncate(text, limit=4096) helper; _handle_question now sends _truncate(answer) to stay within Telegram's 4096-character hard limit.

- rag_client: move httpx.AsyncClient instantiation outside the tenacity retry loop so the connection pool is reused across attempts rather than torn down and rebuilt on every retry - discord_bot: prefer payload.member (populated directly by the gateway event) over a cache lookup in on_raw_reaction_add; fall back to guild.get_member() only when payload.member is None - telegram_bot: add _truncate() helper (4096-char Telegram API limit) and apply it in _handle_question() to prevent MessageTooLong errors Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

0xPepeSilvia · 2026-04-15T08:52:41Z

All three points from the gemini review have been addressed in 99fab49:

1. HTTP client instantiation outside retry loop (rag_client.py)
httpx.AsyncClient is now created once before the AsyncRetrying loop — the connection pool is reused across retry attempts rather than torn down and rebuilt on every attempt.

2. Discord member from event payload (discord_bot.py)
on_raw_reaction_add now reads payload.member first (populated directly by the Discord gateway event) and only falls back to guild.get_member() when payload.member is None. This avoids false negatives from an incomplete member cache.

3. Telegram message length truncation (telegram_bot.py)
Added a _truncate(text, limit=4096) helper matching the pattern already in the Discord bot. _handle_question now sends _truncate(answer) to stay within Telegram's Bot API hard limit and prevent MessageTooLong errors.

0xPepeSilvia added 6 commits April 15, 2026 05:58

feat: add docker-compose.yml orchestrating AnythingLLM + bridge

00d4c8f

feat: add .env.example with all required configuration

d951b42

docs: add ARCHITECTURE.md with framework selection rationale

f772c4a

docs: rewrite README with full setup guide

b1555d7

feat: add multi-channel AI bridge (Telegram + Discord + feedback)

7b4efbd

chore: update .gitignore to exclude pycache and session files

12c25b5

gemini-code-assist Bot reviewed Apr 14, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: replace faqqer with multi-channel AI help bot (AnythingLLM + Discord + Telegram)#3

feat: replace faqqer with multi-channel AI help bot (AnythingLLM + Discord + Telegram)#3
0xPepeSilvia wants to merge 7 commits into
tari-project:mainfrom
0xPepeSilvia:feat/ai-help-bot-v2

0xPepeSilvia commented Apr 14, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Uh oh!

0xPepeSilvia Apr 16, 2026

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Uh oh!

0xPepeSilvia Apr 16, 2026

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Uh oh!

0xPepeSilvia Apr 16, 2026

Uh oh!

0xPepeSilvia commented Apr 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		answer = await self._answer(question)
		sent = await msg.reply_text(answer)

Conversation

0xPepeSilvia commented Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Acceptance criteria

Framework rationale — why AnythingLLM

Architecture

Quick deploy

Test results

Bounty payout

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

0xPepeSilvia Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

0xPepeSilvia Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

0xPepeSilvia Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

0xPepeSilvia commented Apr 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

0xPepeSilvia commented Apr 14, 2026 •

edited

Loading