f4bfc12293
Hybrid without a reranker keeps each database's vector and FTS branches apart and compares every candidate by per-branch z-score, summing both branches for a chunk found in each. Databases are then compared by how exceptional a hit is for them rather than by raw score, which is not comparable across indexes. Targets the measured ceiling on rank-and-score fusion: 73% of candidates at n=4 and 84% at n=8 tie on both score and rank, so no key built from those two can separate them and they fall to declaration order under every other arm. Continuous keys should barely collide. Implementation from the multi-fusion session; branch depth via HAIKU_RAG_BRANCH_DEPTH, default 20. Claude-Session: https://claude.ai/code/session_01WhudUtZm6qqiuv8Y1sbwSc |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| conftest.py | ||
| helpers.py | ||
| test_capabilities.py | ||
| test_citations.py | ||
| test_documents.py | ||
| test_expansion.py | ||
| test_lifecycle.py | ||
| test_scope.py | ||
| test_search.py | ||