Retrieval feeds the Enrichment prompt (#125)

[wschenk]

Summary

Research stops starting cold. Before a report is written, what this walk already worked out about the same things is retrieved into the prompt — Threads that named the same nouns first, then ones that merely read alike — each with the reason it is there, and a plain instruction to build on them and say which one.

The chicken-and-egg this had to solve. The Thread being enriched has neither mentions nor an embedding of its own: it earns both from the report about to be written. So both retrieval paths start from the Capture itself.

• The embedding fallback embeds the walker's words at query time and asks the index with that vector, rather than looking the Thread up and finding nothing.
• Mention-first — which would otherwise never fire on a first Capture, the common case — matches nouns the corpus already knows against the words the walker just used. A known noun appearing whole in what they said is the same exact link a stored mention would have been. Names under four characters are skipped: a short noun matches everything and means nothing.

Best-effort by construction. A Capture with no history is prompted exactly as it was before any of this existed. A retrieval that throws — a broken index, a gateway that will not embed — writes the cold-start report rather than failing the job. Every failure path is caught and returns an empty list.

Closes #125

Test plan

• tests/enrichment-retrieval.spec.ts — a Capture whose past shares a noun gets that Thread in its prompt, with the walk date and the shared noun named; a Capture with no history produces a prompt containing no retrieval section at all; a repository whose index throws still completes with the cold-start report; a Thread linked only by resemblance is retrieved and labelled "reads alike"; the renderer's exact output; and the text matcher (whole words only, short names skipped, case-insensitive)
• pnpm typecheck clean; pnpm test — 337 passed, 1 failed (trash-repository.spec.ts, Neon-backed, fails identically on a clean checkout of main here), 10 skipped
• pnpm lint — 6 errors, all pre-existing

Two tests hardened. desk-priors-desktop failed once under full-suite load and passed on re-run. The cause is real and mine: rows paint from the local store before their Enrichments land, so asserting a mention-derived chip without first waiting for a mention-derived element is a race that only shows when both browser projects run at once. Both assertions now gate on the Mentions rail row, which cannot exist until Enrichments have loaded. I ran the full suite twice after the change.

Note for production

The retrieval quality depends on embeddings existing. Until mise run embeddings:backfill is run (after checking the model against the gateway's live /v1/models table), only Threads enriched since #124 have vectors — so the fallback half is thin and mention matching carries it. That is the intended degradation, not a defect.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
walking-thoughts	Ready	Preview	Aug 4, 2026 1:37pm