mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-30 19:49:14 +00:00
85286a556c
* fix(search): stop boosting compiled_truth at default detail (#3430) COMPILED_TRUTH_BOOST = 2.0 is applied AFTER RRF normalization, and RRF's whole dynamic range over a 100-deep pool is 1/60 -> 1/160 (a factor of 2.67). So a 2.0x multiplier consumes roughly three quarters of the range: break-even is `2/(60+r) >= 1/60`, i.e. r <= 60, which means ANY boosted chunk inside the first 60 ranks outranks an unboosted rank-1 chunk. That is a categorical filter, not a tilt. Measured against master's own rrfFusion, with the correct answer in a fenced_code chunk at vector rank 0: compiled_truth chunks in pool | final rank | in top-20 10 | 10 | yes 20 | 20 | NO 40 | 40 | NO 80 | 59 | NO With the boost off the answer stays at rank 0 in every case. The gate was spelled `detail !== 'high'` -- written as though `high` were the special case. The documented contract in src/core/operations.ts is "low (compiled truth only), medium (default, all with dedup), high (all chunks)", which makes LOW the special one: `low` already restricts to compiled_truth, so a boost there is a no-op among equals, while `medium` and `high` are both meant to see everything. So the default detail was silently compiled-truth-only, contradicting the op's own description. Three changes: 1. The three fusion call sites now route through a named predicate, `shouldBoostCompiledTruth(detail)`, returning true only for 'low'. Extracted rather than left inline precisely because an inline expression is only reachable through a full hybridSearch round trip -- which is why the inversion went unnoticed. The predicate is directly unit-testable. 2. KNOBS_HASH_VERSION 13 -> 14. Results are cached AFTER fusion, so rows ranked under the old semantics would otherwise be served under the new ones for the whole TTL (3600s default). One-time miss spike on upgrade. 3. test/search-compiled-truth-boost-scope.test.ts pins both the mapping and the arithmetic, and documents the displacement it prevents. Verified the tests discriminate: stubbing the OLD predicate body into master (so the failure is behavioral rather than a missing export) gives 4 fail / 3 pass; with the fix, 7 pass. typecheck clean, verify 32/32, and 144 pass / 0 fail across the search + fusion + cache suites. * fix(test): update the three remaining KNOBS_HASH_VERSION pins to 14 (#3430) Missed in the first pass because I ran a targeted set of test files instead of the full suite. CI shards 3, 8 and 10 caught them: test/search/knobs-hash-reranker.test.ts:67 test/cross-modal-phase1.test.ts:139,149 test/search-alias-resolved-boost.test.ts:93 Each carries the running history of why the version moved, so each gets the 13→14 rationale appended rather than just the number swapped. No pins at 13 remain anywhere in test/. --------- Co-authored-by: Garry Tan <garrytan@gmail.com>