Repository navigation
P1: validate HNSW retrieval on representative larger corpus #27
Description
The pgvector HNSW index exists and the read-only KNN diagnostic recovered 20/20 exact top candidates from 70 on the small current corpus, but PostgreSQL chose a sequential scan. Acceptance: (1) isolated, clearly synthetic benchmark with at least 10k vectors without modifying production observations; (2) EXPLAIN query-plan evidence for ANN vs exact; (3) recall@K measured over current-source and independent-root eligibility, with missed-pair review; (4) p50/p95 latency and memory/cost bounds; (5) do not switch the production query unless recall targets and resource limits are met; (6) CI-safe deterministic tests. Never claim synthetic benchmark measures real-world corroboration accuracy.
Activity
guillaume-flambard commentedon Oct 12, 2026
P1 scalability progress: PR #29 adds isolated 10k-vector deterministic synthetic benchmark. Local PostgreSQL EXPLAIN used the HNSW index. Across six synthetic anchors, top-10 recall vs exact search was 10/10 for each; ANN median/p95 0.66/1.47 ms, exact median/p95 5.50/17.19 ms. All measurements are synthetic, limited to six anchors and include DB round trips. Production candidate retrieval is NOT changed. 263 PostgreSQL-inclusive tests pass. Still needed before closing: representative real-source corpus with enough source-grounded paraphrases, recall and p95 measurements across independent threads, and production-scale index-plan proof. The current 44-source corpus is far too small. Do not interpret this as real-world semantic accuracy or proof of production performance.