Amr Gharbeia 11254b56ec v0.4.0: semantic retrieval activation — wire foveal-vector + trigram Jaccard ...

1. Wire :foveal-vector into context-awareness-assemble: pass the foveal
   node's embedding vector to context-object-render. Previously always
   nil → similarity always 0.0 → no semantic boosting.

2. Replace default :hashing (SHA-256) with :trigram (character-trigram
   Jaccard). SHA-256 is a cryptographic hash with the avalanche property
   — one-bit input differences produce entirely different outputs. Useless
   for similarity. Trigram bloom filter (128-dim) captures lexical overlap
   in pure Lisp with zero external dependencies:
   - 'authentication' vs 'authenticate' → 0.80 similarity
   - 'authentication' vs 'banana' → 0.00 similarity

3. Rename old embedding-backend-hashing → embedding-backend-sha256
   (integrity-only, explicit opt-in). Add embedding-backend-trigram.

4. Add test-semantic-retrieval-trigram: related texts > 0.75, unrelated < 0.3.

Test: 97/0 across 13 suites (context 12/0, embedding 12/0)

2026-05-06 19:04:17 -04:00

11 KiB

Raw Blame History

View Raw