Crabhaus

What it is

Hybrid search runs a keyword engine and a vector engine over the same corpus and merges their ranked lists, so one query gets exact term matching and semantic similarity at once. A reranker, usually a cross-encoder, then re-scores the merged shortlist by reading the query and each candidate document together before the final ordering.

How it works

BM25 scores documents on distinctive exact-term matches; an embedding index retrieves semantic neighbors that may share no words with the query. Reciprocal Rank Fusion merges the two lists, giving each document 1/(k + rank) per list, so no score calibration is needed. A cross-encoder then reads each query-document pair jointly, an accurate but slow pass reserved for the fused top 50-100.

A concrete example

Searching internal docs for 'reset MFA token for contractor accounts': BM25 nails pages containing the literal strings MFA and contractor, vector search surfaces a runbook titled 're-enrolling two-factor devices for external staff', RRF merges both lists, and the reranker reads the actual passages and puts that runbook first.

Why it matters

Retrieval quality caps RAG and agent-memory quality. Embedding-only search fails on identifiers, error codes, and rare names; keyword-only misses paraphrase. BEIR showed dense retrievers often lose to plain BM25 out of domain while cross-encoders rank near the top, so hybrid plus rerank is a high-leverage upgrade far cheaper than a bigger generation model.