0549beb156
The README table was a pair of unexplained numbers. Spell out recall vs ranking here, and expand EVAL.md §2 with the formulas used on this corpus. Co-authored-by: Cursor <cursoragent@cursor.com>
123 lines
5 KiB
Markdown
123 lines
5 KiB
Markdown
# depiedencap-rag — Discourse-embedded RAG agent
|
|
|
|
📘 Ce projet est également disponible en [français 🇫🇷](./README.fr.md)
|
|
|
|
> ⚠️ **Work in progress.** The agent is currently being tested on a
|
|
> restricted demo group and is **not publicly deployed** on the forum.
|
|
> This repository documents the work — no demo URL is published.
|
|
|
|
A **Discourse plugin** + a self-hosted RAG pipeline, built for a French
|
|
community forum (Depiedencap — men's shoes, 424 835 posts / 22 230
|
|
topics). Product principle: **the bot points members to existing threads
|
|
instead of answering in their place.**
|
|
|
|
## What it does
|
|
|
|
- **Post-grain retrieval** (not topic-grain): `bge-m3` embeddings in
|
|
`ai_posts_embeddings` (pgvector inside the forum's Postgres on the VPS),
|
|
top-25 by cosine then **cross-encoder rerank** (TEI on the mini PC),
|
|
per-thread dedup.
|
|
- **Multi-turn query rewriting**: the latest message plus earlier turns
|
|
are condensed into a standalone query (*condense question* /
|
|
history-aware retriever). Same resident model as generation:
|
|
`qwen3:30b-a3b-q6k` (MoE, ~3B active, `think:false`) — not a separate
|
|
small rewriter.
|
|
- **Server-side verified citations**: the LLM only emits bare `[[n]]`
|
|
markers; the plugin maps them to `/t/slug/id/post_number` links and
|
|
discards any link that does not actually exist in Postgres (URL
|
|
whitelist — the LLM never gets the last word on a link).
|
|
- **Graduated evidence zones** (abstain / cautious / normal) driven by
|
|
reranker score + a **deterministic input gate** (intent regexes) for
|
|
dangerous or out-of-role requests — zero LLM calls.
|
|
- Graceful degradation: embedding or reranker endpoint down → legacy
|
|
fallback / honest abstention, never invention.
|
|
|
|
## Architecture
|
|
|
|
The plugin runs **inside Discourse on the VPS**. The mini PC is only the
|
|
inference backend (Ollama + TEI), reached over Tailscale. Gate, pgvector
|
|
and the citation sanitizer never leave the VPS.
|
|
|
|
```
|
|
Member question
|
|
│
|
|
▼
|
|
[VPS] Discourse + this Ruby plugin
|
|
│
|
|
├─ 1. gate regex, local, 0 LLM
|
|
│
|
|
├─ 2. rewrite ──Tailscale──► [mini PC] Ollama qwen3:30b-a3b-q6k
|
|
│
|
|
├─ 3. embed ──Tailscale──► [mini PC] Ollama bge-m3
|
|
│
|
|
├─ 4. pgvector Postgres local, 424 835 posts, cosine top-25
|
|
│
|
|
├─ 5. rerank ──Tailscale──► [mini PC] TEI mmarco-mMiniLMv2
|
|
│
|
|
├─ 6. generate ──Tailscale──► [mini PC] Ollama qwen3:30b-a3b-q6k
|
|
│
|
|
└─ 7. sanitizer URL whitelist, local Postgres
|
|
│
|
|
▼
|
|
answer + /t/slug/id/post_number links
|
|
```
|
|
|
|
| Lives on the VPS | Lives on the mini PC (k3s, iGPU) |
|
|
|---|---|
|
|
| Discourse, this plugin, Postgres/pgvector | Ollama (`bge-m3` + `qwen3:30b-a3b-q6k`) and TEI (`mmarco`) |
|
|
|
|
`qwen3:30b-a3b-q6k` is a single resident MoE (~3B active) used for both
|
|
rewrite (`think:false`) and generation.
|
|
|
|
Deliberate choice: **no RAG framework** (LangChain & co.) — a linear
|
|
pipeline in direct code, instrumented for evaluation. Full reasoning:
|
|
[`docs/rag/DESIGN.md` §9](docs/rag/DESIGN.md#user-content-9-pourquoi-pas-langchain).
|
|
|
|
## Measured results (real corpus, 812-question eval set)
|
|
|
|
| Metric | Value |
|
|
|---|---|
|
|
| Hit@5 retrieval | **0.97** |
|
|
| MRR | 0.89 |
|
|
| Safety (red-team, gate, refusal, evidence injection) | 20/20 |
|
|
| Rerank overhead | ~100 ms |
|
|
|
|
**Hit@5** : share of eval questions whose expected thread (or post) appears
|
|
in the **top 5** retrieved results. 0.97 means the right source is in that
|
|
shortlist 97 times out of 100. It does not say *where* in the five.
|
|
|
|
**MRR** (Mean Reciprocal Rank) : scores the **rank of the first** relevant
|
|
hit, then averages. Rank 1 → 1, rank 2 → 0.5, rank 5 → 0.2, not in the
|
|
list → 0. 0.89 means the first good hit is usually 1st or 2nd, not merely
|
|
"somewhere in the top 5".
|
|
|
|
Hit@5 = recall ("did we miss it?"). MRR = ranking ("did we put it first?").
|
|
Protocol and formulas :
|
|
[`docs/rag/EVAL.md` §2](docs/rag/EVAL.md#user-content-2-métriques).
|
|
|
|
Details: [`docs/rag/EVAL.md`](docs/rag/EVAL.md) ·
|
|
[`docs/rag/SAFETY.md`](docs/rag/SAFETY.md) ·
|
|
[`docs/rag/PIPELINE.md`](docs/rag/PIPELINE.md) ·
|
|
[`docs/rag/REFERENCES.md`](docs/rag/REFERENCES.md)
|
|
|
|
## Repository layout
|
|
|
|
| Path | Content |
|
|
|---|---|
|
|
| root (`plugin.rb`, `lib/`, `app/`, `config/`) | The Discourse plugin — installable via `git clone` into `plugins/` |
|
|
| `docs/rag/` | Design, evaluation, safety, references, local dev recipe |
|
|
| `prompts/` | Production system prompt (v3.1) |
|
|
| `tools/` | Live smoke test (Discourse API) |
|
|
|
|
## Security & privacy
|
|
|
|
- No internal endpoint, IP or account is published (sanitised export from
|
|
the private working repo + automated denylist check).
|
|
- The agent stays limited to a demo group during the observation phase;
|
|
documented < 5 min rollback.
|
|
- The bot refuses dangerous, illegal or generalist requests and ignores
|
|
hostile instructions found inside retrieved forum excerpts.
|
|
|
|
## License
|
|
|
|
AGPL-3.0 — see `LICENSE`.
|