feat!: P0 iron-out — generic deletes, de-fashioned core, honest ESCI judge, one env contract - #87
Merged
Merged
Conversation
35+ session scratchpads / implementation notes / audit artifacts moved out of the repo root so it presents as a product repo, not a build log. Includes the iron-out session's own notes (assumptions, root causes, verification record). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- system-behavior-spec.md: evidence-grounded spec of the 2.6.0 system (indexing spine, enrichment, hybrid retrieval, eval, tenancy truth). - stage-fit-audit-and-iron-out-plan.md: keep/kill verdicts on every abstraction + prioritized P0-P2 backlog toward the marketplace-first, replaceable-search direction; executed-work record. - docs/research/mices/: 11 MICES talks (idealo, Zalando, dm, Digitec, Delivery Hero, Shopify, OTTO, MediaMarkt, Coveo, Qdrant, Turnbull) synthesized with bet-by-bet verdicts for samesake. - Commits the previously untracked research corpus (doordash, mastra, qmd, conversational-commerce-search, postgres-high-scale-search, ...), product audits, and RFC drafts the plan references. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… indexing path
Breaking (see .changeset/tier-zero-defaults.md; recreate + reindex collections;
requires pgvector >= 0.7, 0.8 recommended):
- Collection embedding/space_vec columns and HNSW indexes are halfvec (fp16):
measured 2.25x smaller index, 3.6x faster build, zero relevance delta on the
62-query golden eval (grade@5 1.878 == baseline, nDCG@5 0.902 vs 0.901).
Dim ceiling 4000 for collections; entity tables stay vector (2000).
- Weighted lexical leg: fts = setweight(fts_src_a,'A') || setweight(fts_src,'B');
opt-in via f.text({ searchable: true, ftsWeight: "A" }) or an fts indexing
surface with weight "A". Dead CollectionTextFieldDef.weight removed.
- Deprecated DEFAULT_PRODUCT_PARSE_INSTRUCTIONS removed (use _BODY).
- Legacy fashion preset layer deleted (fashionSearchPreset & co — zero callers).
Fixed:
- Since the S1c indexing migration, collections without an enrich pipeline
indexed nothing (surfaces only built during enrich) — hello-search, quickstart,
and the README example were silently broken. `indexing` is optional again,
CollectionEmbeddingDef.source is restored, and embed-index builds surfaces
inline (declared surfaces win; enrich-owning collections unchanged).
- README version claim (1.0.0 -> 2.6.0) and pgvector requirement noted.
Added:
- Iterative index scans (hnsw.iterative_scan = relaxed_order) auto-enabled on
vector legs when pgvector >= 0.8 — filtered recall went 2/100 -> 37/100 rows
on a 0.5%-selective filter in the Neon benchmark.
- efSearch (10-1000) search option, in-process + HTTP, in the result-cache key.
- StorageAdapter.pgvectorVersion() + unsafeWithSettings() (SET LOCAL-scoped
queries); apply fails fast when pgvector < 0.7.
Verified: tsc clean; server suite 261/261; sdk 4/4; hello-search, hello-spaces,
quickstart release gates pass; golden eval flat-or-better vs July 1 baseline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds the June/July eval artifacts (fashion judge v1, enrichment fixture/live, search p2post baseline) and the 2026-07-02 tier0post run proving the halfvec + iterative-scans change is relevance-neutral: mean grade@5 1.878 (== p2post baseline), nDCG@5 0.902 (+0.001), no-results 0%, typo nDCG +0.005. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… judge, one env contract
P0-1 removeDocuments on every surface: HTTP DELETE
/v1/projects/:p/collections/:c/documents, CLI `samesake remove`, and
catalog-sync deletes rerouted through it. Proof test: push -> delete ->
search returns nothing (HTTP + in-process); CLI proven against a live server.
P0-3 De-fashioned the generic core: core/fashion-search.ts replaced by
vertical-neutral core/shop-search.ts (shopSearch, /shop-search) and
core/catalog-sync.ts (syncCatalogEvent, /catalog-sync). Zero-result filter
relaxation is now opt-in via CollectionSearchDef.relaxableFilters (core
relaxes nothing by default; fashion template ships fashion.searchDefaults()).
SDK renames: ShopSearchRequest/Response, ShopperContext, CatalogSyncEvent;
FashionRankingPolicy deleted; fashionRerank -> llmRerank. Eval constraints
now use the search filter vocabulary ({price: {$lte: N}}) resolved against
the schema. Grep gate: defashion-gate.test.ts forbids fashion symbols in core.
P0-4 Judge honesty: shared 4-class ESCI rubric (E=3/S=2 soft positive/C=1/I=0,
floor 2) across makeLlmJudge and evaluateSearch; judge version content-hashed
(<tag>@<sha256(rubric)[:8]>) so prompt edits invalidate caches; same-family
enrich+judge rejected (runEval, evaluateSearch, calibrateSearch, HTTP
judgeModel). New honest baseline minted with a cross-family gpt-4.1-mini
judge: evals/runs/2026-07-02T16-01-22-852Z-search-p0honesty.json — mean
grade@5 1.881 / nDCG@5 0.901 vs tier0post 1.878/0.902, topIds 61/62
identical (one embedding-jitter reorder).
P0-5 One env contract: SAMESAKE_DATABASE_URL / SAMESAKE_API_KEY canonical across
.env.example, README, docs, examples, apps, tests, CLI; apps/matcher env
shim deleted; provider keys provider-named (GEMINI_API_KEY, OPENAI_API_KEY);
GOOGLE_GENERATIVE_AI_API_KEY eliminated. No fallback aliases.
Verified: tsc --noEmit clean; server suite 265 pass / 0 fail (baseline 258);
hello-search, hello-spaces, quickstart green post-rebuild; docs site builds.
Breaking changes folded into the pending tier-zero-defaults major changeset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The four P0 correctness-and-honesty items from
docs/stage-fit-audit-and-iron-out-plan.md§3, shipped as one breaking change (folded into the pendingtier-zero-defaultsmajor — merging this and the follow-up version PR releases it):removeDocumentson every surface. HTTPDELETE /v1/projects/:p/collections/:c/documents, CLIsamesake remove --ids=…, and catalog-sync deletes rerouted through the same path. Proof: push → delete → search returns nothing (HTTP + in-process test), plus a live CLI run against a real server.core/fashion-search.ts→ vertical-neutralcore/shop-search.ts(shopSearch,/shop-search) +core/catalog-sync.ts(syncCatalogEvent,/catalog-sync). Zero-result relaxation is opt-in viaCollectionSearchDef.relaxableFilters(core relaxes nothing by default; the fashion template shipsfashion.searchDefaults()). Eval constraints use the search filter vocabulary ({price: {$lte: N}}).defashion-gate.test.tsgreps core for fashion symbols and fails on any.<tag>@<sha256(rubric)[:8]>) so prompt edits auto-invalidate caches; same-family enrich+judge throws (runEval,evaluateSearch,calibrateSearch, HTTPjudgeModel).SAMESAKE_DATABASE_URL/SAMESAKE_API_KEYcanonical everywhere; matcher env shim deleted; provider keys provider-named;GOOGLE_GENERATIVE_AI_API_KEYeliminated. No fallback aliases.Golden eval — new honest baseline
Cross-family judge (gpt-4.1-mini over the Gemini-enriched corpus), all 62 queries:
evals/runs/2026-07-02T16-01-22-852Z-search-p0honesty.json— mean grade@5 1.881 · nDCG@5 0.901 · no-results 0% (tier0post: 1.878 / 0.902). Retrieval flatness: 61/62 identical per-query topIds; the one delta is embedding float jitter on a pure-semantic query (query embeddings recompute live per run), not a retrieval change.Verification
tsc --noEmitclean across workspacespackages/serversuite: 265 pass / 0 fail (baseline 258)relaxableFilters, env contractdocs/notes/p0-iron-out-implementation-notes.md🤖 Generated with Claude Code