CONCEPT Cited by 6 sources

Hybrid Search¶

Definition¶

Hybrid search combines lexical retrieval (BM25 / keyword / exact-term) with semantic retrieval (dense vector similarity) — usually running both in parallel and fusing their result lists — so that a single query gets both the exact-term precision of keyword search and the paraphrase/synonym recall of vector search. It is the dominant production retrieval shape for RAG and agentic search, precisely because neither mode alone is sufficient:

Lexical (BM25) nails acronyms, proper nouns, IDs, and specific error strings — cases where embedding similarity drifts.
Semantic (vectors) covers paraphrase ("bought a bicycle" ≈ "purchased a bike") — cases BM25's exact-token matching misses.

See concepts/retrieval-ranking-funnel for how hybrid retrieval sits inside a broader retrieve → fuse → rerank funnel.

Fusion methods¶

The two ranked lists are combined by one of:

Reciprocal Rank Fusion (RRF) — score each doc by Σ 1/(k + rank_i) across lists; robust, no score calibration needed.
Weighted score sum — normalize and blend the raw scores (requires calibrating lexical vs vector score scales).
Learned reranker — a cross-encoder or gradient-boosted model re-scores the union of candidates. See patterns/parallel-retrieval-fusion.

Why it is a first-class production requirement¶

Multiple production systems keep BM25 as a primary surface rather than a legacy fallback, and layer vectors alongside it:

Dropbox Dash runs a BM25 lexical index and dense vectors in a hybrid index, calling BM25 "an amazing workhorse" effective on its own. (Source: sources/2026-01-28-dropbox-knowledge-graphs-mcp-dspy-dash)
Cloudflare AI Search promotes hybrid to a config knob: index_method: { keyword: true, vector: true }, with tokenizer and match-mode options. Its GA query lifecycle runs vector + keyword search in parallel, fuses, then optionally reranks — and now accepts image queries on the vector half (embedded directly by a multimodal model, or captioned first for text-only models). (Source: sources/2026-04-16-cloudflare-ai-search-the-search-primitive-for-your-agents; sources/2026-10-01-cloudflare-ai-search-is-now-generally-available)
Expedia's Embedding Store exposes similarity search combined with attribute/metadata predicates (price < 100, category = electronics) as its two query surfaces. (Source: sources/2026-01-06-expedia-powering-vector-embedding-capabilities)

Native hybrid search inside the database¶

Historically hybrid search meant stitching a dedicated search engine to the primary database with an ETL pipeline. Databricks' Lakebase Search collapses this into Postgres itself: combining lakebase_text (BM25) with lakebase_vector (ANN) lets a single SQL query fuse semantic + keyword relevance, apply SQL filter predicates, and join directly against live operational tables — "one SQL tool call for agents" replacing a multi-system retrieval pipeline. (Source: sources/2026-09-28-databricks-lakebase-search) This is the in-database realization of hybrid search: the fusion happens where the operational data already lives, governed by standard database rules.

Seen in¶

sources/2026-10-02-databricks-real-time-retail-intelligence-building-e-commerce-recommenda — "hybrid" as ANN + hard metadata filters in a recsys candidate stage. Databricks' retail reference architecture describes AI Search performing "hybrid retrieval: ANN similarity search combined with hard metadata filters (category eligibility, regional inventory, minimum stock thresholds)." Note this is a different sense of "hybrid" than lexical+dense (BM25+vector) fusion — here it's semantic similarity + structured eligibility predicates applied inside retrieval, so ineligible/out-of-stock items never enter the candidate set. Reinforces that production "hybrid search" often means fusing vector relevance with hard business/inventory constraints, not just two relevance signals.
sources/2026-09-28-databricks-lakebase-search — native hybrid search in Postgres via lakebase_text + lakebase_vector fused with SQL filters and joins; Conexiom runs BM25 hybrid over 100M+ rows.
sources/2026-01-28-dropbox-knowledge-graphs-mcp-dspy-dash — BM25 + dense vectors as a hybrid index; knowledge-graph signals layered on top as ranking input.
sources/2026-04-16-cloudflare-ai-search-the-search-primitive-for-your-agents — hybrid search as a managed config knob with keyword + vector methods.
sources/2026-10-01-cloudflare-ai-search-is-now-generally-available — GA query lifecycle (parallel vector + keyword → fuse → rerank) plus multimodal image queries on the vector half.
sources/2026-01-06-expedia-powering-vector-embedding-capabilities — similarity search combined with attribute/metadata predicates.

systems/bm25 — the lexical half.
concepts/vector-similarity-search — the semantic half.
concepts/retrieval-ranking-funnel — the broader funnel.
patterns/parallel-retrieval-fusion — parallel retrieve + fuse pattern.