[wufangyong973]

Closes #38.

New contributors had to reverse-engineer how retrieval works from the code and AGENTS.md. This adds a short docs/ARCHITECTURE.md that lowers that barrier, covering the scope called out in the issue:

• The retrieval pipeline — query → three-tier FTS (AND / OR / stemmed prefix) → vector search → entity matching → weighted RRF fusion, with where each leg lives (fts.ts, vector-index.ts + embeddings.ts, entities.ts, fusion.ts, assembled in context.ts).
• Why the FTS is three tiers, and why the IDF is applied on the query side only (so stored vectors stay a pure function of their own text).
• The migration model — sequential named, append-only migrations in migrate.ts.
• The thin-tools convention — logic in lib/, tools just wire it up.
• The working agreement — measure retrieval changes with npm run bench, and the local-first no-LLM / no-network constraint.

Every claim is grounded in the existing source and its module header comments (e.g. the query-side-IDF invariant in idf.ts, the three-tier design and the stemmed-prefix +9.1 hit@5 gain in fts.ts, the weighted-RRF and re-sweep rule in fusion.ts).

Wiring: link it from README.md, and drop the now-stale "(if present)" hedge from the docs/ARCHITECTURE.md link in CONTRIBUTING.md.

[vercel]

@wufangyong973 is attempting to deploy a commit to the vignesh's projects Team on Vercel.

A member of the Team first needs to authorize it.