Repository navigation
feat: zero-download semantic fallback — algorithmic embeddings + code-aware FTS tokenizer #169
Description
From the codebase-memory-mcp analysis (2026-07-25): they ship semantic search with ZERO model download — 11-signal algorithmic embeddings (TF-IDF, Random Indexing with co-occurrence synonym bridging, MinHash, API/type signatures, AST profile, graph diffusion, Halstead-lite; verified src/semantic/semantic.h) optionally seeded by static int8 token vectors compiled into the binary. No sentence-transformers, no Ollama, no first-run cliff.
Cortex's SQLite-mode installs currently degrade until sentence-transformers finishes installing (session-start note: 'semantic search will improve next session'). Proposal: an algorithmic-embedding fallback engaged automatically when the neural model is absent — same 384-dim space contract, marked lower-fidelity in telemetry, upgraded transparently when the model arrives. Plus: camelCase/snake_case-aware FTS tokenization for code-flavored memories (their cbm_camel_split FTS5 tokenizer).
Zetetic gate: benchmark the fallback on LongMemEval-S/LoCoMo vs (a) no-vector baseline and (b) sentence-transformers — publish the three-way numbers; adopt only if fallback beats no-vector materially. Done criteria: fresh SQLite install answers a semantic recall correctly with zero network; benchmark table committed.
Activity
cdeust commentedon Jul 24, 2026
Implementation delivered: PR #172 (
d5835c0). Three-way benchmark committed (LongMemEval-S, SQLite, n=50): no-vector 0.275 MRR / 46.0% R@10 → algorithmic fallback 0.378 / 66.0% → sentence-transformers 0.609 / 94.0%. Adoption criterion met (+0.102 MRR, +20pp R@10 over no-vector). Constants sourced (CBM semantic.h / Manning IR / Mikolov 2013 / Achlioptas 2003); mixed-store no-cross-rank invariant tested; telemetry marks fallback mode. Reported deviations accepted as justified: dedicated SQLite bench harness reusing the existing loader/scorer verbatim (PG harness cannot express the 3-way), and two pre-existing §13 items now filed as dated issues (embedding_engine split; test_rebuild_profiles live-data dependency). Merge on CI green.cdeust commentedon Jul 24, 2026
PR #172 denied (see closing comment). Plan of record: Phase A refactor PR (seams, closes #173) → Phase B full #169 on the refactored base. The n=50 three-way benchmark result (no-vector 0.275/46.0% → fallback 0.378/66.0% → sentence-transformers 0.609/94.0%) remains valid evidence for the approach and carries into Phase B.
cdeust commentedon Jul 24, 2026
Phase A delivered: PR #175 (three seams: EmbeddingProvider protocol, ModelState lifecycle machine, composition-root factory; closes #173). Key finding from the seam work: PR #172's empty-stamp bug was in the SQLite store's _has_vec-gated stamp write, not the engine — the boundary now has an explicit seam and Phase B maps every lifecycle state to a provider so no state can leave an empty stamp. Behavior preservation proven: 56-test embedding suite unchanged, full suite green except the 3 documented #174 failures. Merge on CI green, then Phase B.
Merge pull request #181 from cdeust/feat/semantic-fallback-169b
cdeust commentedon Jul 25, 2026
COMPLETE — delivered across two PRs per the refactor-first rule: PR #175 (Phase A: EmbeddingProvider protocol, explicit ModelState lifecycle, factory seams; closed #173) and PR #181 (Phase B, merged 0db6b2e with 14/14 checks): total ModelState→provider map (no state can leave an empty stamp — the PR #172 denial's root cause is structurally unrepresentable), unconditional SQLite embedding_model stamp, loud engagement + telemetry embedding_mode, camelCase/snake_case tokenization on ALL recall paths, scheduled fallback→neural upgrade in consolidate with counts, genuine-absence subprocess integration test (import-hook, zero network), mixed-store no-cross-rank invariant. Benchmark (LongMemEval-S, SQLite, n=50, MANIFEST committed): no-vector 0.275 MRR/46% R@10 → algorithmic fallback 0.378/66% (+0.102/+20pp, adoption criterion met) → sentence-transformers 0.609/94%. Zero-download semantic search is live for SQLite-mode installs.
feat(embedding): zero-download algorithmic fallback on the cdeust#173 …
bench(embedding): three-way SQLite LongMemEval-S for the cdeust#169 f…
test(embedding): update I2 canonical-writer line pins for cdeust#169 …
fix(embedding): make frequent-token subsampling match its cited CBM s…
fix(sqlite): stamp 'neural' only after the vector actually lands
fix(sqlite): downgrade to 'fallback' instead of skipping the stamp
fix(sqlite): install sqlite-vec on every backend, self-heal vectorles…