Hybrid Retrieval

Run a keyword search and a semantic search in parallel over the same corpus, then fuse the two ranked lists into one. The move exists because each channel has the other's blind spot: Semantic Retrieval (vector/embedding) matches by meaning but is unreliable on literal strings and exact identifiers — an error code like error: connection refused, or a function name like useState, may not surface at all, because embeddings privilege sense over surface tokens. Keyword search (BM25 over an FTS index) is the inverse: exact on tokens, blind to paraphrase. Neither alone covers a memory store whose queries mix natural language with precise identifiers, so the practical answer is both, combined by Rank Fusion. In OpenClaw the weighting is ~0.7 vector / 0.3 keyword.

The corrective this concept carries into the Retrieval Maturity Levels ladder is that "just add semantic search" is not the terminal upgrade — semantic has a shape of failure (exact matches), and hybrid is the least-bad general default precisely when you can't guarantee the query shape in advance. It also composes downstream with Reranking (fuse first, then re-score the shortlist with a model). Note the counter-current: the Claude Code team started with a vector database and moved to grep + agentic search because it performed better and was easier to maintain — evidence that even the keyword half alone can beat the semantic half for a code-shaped corpus. Which channels you actually need is a Query-Shaped Storage decision, not a default.

Claims

Semantic search is unreliable for literal strings and exact identifiers (error codes, function names). principle — durable structural limit: retrieval-by-meaning does not privilege surface tokens, so exact-match queries can miss entirely; it's not a tuning problem.
Keyword and semantic search have complementary blind spots — run both in parallel and fuse the results. best practice — context: a memory/RAG store whose queries mix natural language with precise identifiers; "best" because neither channel alone covers both query shapes. If your queries are reliably one shape, you may not need both.
The Claude Code team started with a vector DB and moved to grep + agentic search — simpler and easier to maintain. observation — an independent report that the keyword/agentic channel can outperform the vector channel for a code-shaped corpus; corroborates "no pain, no climb" (Agentic Simplicity, Retrieval Maturity Levels).
BM25 (what powers FTS5) ranks keyword hits by term frequency × term rarity, not just presence. observation — factual: it's a relevance-ranked keyword search, a step up from plain grep's literal match, and the keyword half of most hybrid setups.

Semantic Retrieval — the meaning-match channel hybrid pairs with keyword; its literal- string weakness is exactly what hybrid patches.
Rank Fusion — the mechanism that merges the two ranked lists (weighted vs RRF).
Reranking — the optional model-scored pass applied after fusion.
Retrieval Maturity Levels — hybrid refines level 3: semantic alone is not the terminus.
Query-Shaped Storage — whether you need both channels is set by the query shape.
Agentic Simplicity — the Claude-Code-dropped-vectors note is "no pain, no climb" again.
Contextual Retrieval — the index-side complement: situate chunks before embedding so both channels hybrid fuses carry document context.
Distillate: How AI Agents Search Their Memory — Hybrid Retrieval, in Practice (OpenClaw)
Distillate: Anthropic's Claude Cookbooks — the canonical recipe index — RAG + contextual-embeddings recipes.

Hybrid Retrieval

Claims

Related

Linked from