Experimental Hybrid Search
Normal Fetch remains Engine A and is the default search. This optional lab compares the same query through A: current Fetch retrieval, B: page/evidence-unit SQLite FTS5/BM25, and C: bounded fusion with local hierarchical semantic reranking. B and C use isolated evidence units before the unchanged Fetch verifier and Phase 6G authority sequence. Engine C remains experimental and is not production-approved.
Local-data safeguard
Existing extracted page text is read from this browser and placed only into temporary in-memory indexes. The SQLite runtime, Transformers.js runtime and embedding model are self-hosted by this development build. No query, page text or embedding is sent to a cloud service.
First use loads local static model assets into ordinary browser/service-worker cache. FTS and semantic vectors are not written to IndexedDB or OPFS and disappear when this page closes.
1. Build temporary indexes
Checking the existing Fetch library…
Model: not loaded · proposed baseline Xenova/all-MiniLM-L6-v2
2. Run the same query through all engines
FTS query: —
Build the temporary indexes to begin.
3. Evaluation runner
Runs known-good, historical-failure and new development cases. The locked Set 3 holdout is excluded from the normal evaluation and may be run once after the milestone configuration is frozen.
No evaluation run yet.