Repository navigation
tech-debt(rag): two hybrid-search implementations — one reachable only from a validator #17599
Description
Parent: #10603
HybridSearchEngine (autobot-backend/utils/hybrid_search.py:133, 507 lines) has exactly one
caller: autobot-backend/utils/system_validator.py:894 — a validator, not a production path.
The live RAG path uses a different hybrid implementation, the BM25 route in
autobot-backend/advanced_rag_optimizer.py:475 landed by #10600.
So we carry two implementations of one concept, and the more elaborate one is wired only to a
self-check. Per consolidate, never fork, one of these is unwired machinery.
Verification run 2026-09-27:
grep -rn "get_hybrid_search_engine|HybridSearchEngine" --include=*.py .
utils/hybrid_search.py:133,496,503 (definition)
utils/system_validator.py:894,910 (only caller)
repo_tests/kb_read_visibility_allowlist.py:204-206
Note its keyword scorer is a hand-rolled word-frequency map
(_build_word_frequency_map, line 67) — naive TF, not BM25 — so it is also the weaker of the two.
Acceptance criteria
- One hybrid-search implementation remains, named once, reachable from the production path
- The retained implementation uses BM25, not substring TF
-
system_validatorvalidates the retained implementation, not a private copy - No behaviour is deleted — capability in the dropped copy is carried over or explicitly
recorded as unwanted with a reason
Activity
fix(validator): the hybrid-search check runs the enhanced path, so it…
mrveiss commentedon Oct 12, 2026
Closure evidence (merged in merge train #18358, 13be1b4; constituent #18348 plus a train fix). Verified against
main.HybridSearchEngine(utils/hybrid_search.py) is deleted, after a zero-code-callers check. The canonical path isKnowledgeBase.searchin hybrid mode (HybridSearcher).HybridSearcherfuses BM25 with semantic results (RRF).system_validatorvalidates the retained implementation: this needed a train fix (a060ea5). The original callkb.search(..., top_k=5, mode=HYBRID)took_search_basic, which dropsmode, so the validator never reachedHybridSearcher. It now passeslimit=5, which selects the enhanced path, and readsresultsfrom the dict response (autobot-backend/utils/system_validator.py,validate_hybrid_search). CodeRabbit found the gap on the train PR; the merger verified it againstknowledge/search.py's dispatch.explain_searchandextract_keywords.