Hybrid search — BM25 + vector + reranking

[servathadi]

Context

Current /search uses cosine similarity only. Research shows hybrid retrieval systems cut failure rates from 15% to 5% compared to single-mode vector search.

Motivation

Vector-only search struggles with exact keyword matches, proper nouns, and short queries. BM25 excels precisely where embedding similarity fails. Combining both gives the best of both worlds.

Implementation

Three-stage retrieval pipeline:

1. BM25 full-text search via PostgreSQL tsvector/tsquery on raw_data->>'text'
2. Cosine similarity via pgvector (existing)
3. RRF (Reciprocal Rank Fusion) to merge both ranked lists into a single score: score = 1/(k + rank_bm25) + 1/(k + rank_vector) where k=60

Optional fourth stage: add ColBERT-style reranker for high-stakes queries.

Notes

• pgvector already installed — no new deps needed for BM25 (use PostgreSQL native FTS)
• Add tsvector column to mirror_engrams and populate via trigger
• Add recall_mode=semantic|hybrid|bm25 param to /search endpoint
• Default to hybrid — it's strictly better than semantic alone

References

Zep Memory v2 research benchmarks, pgvector hybrid search docs