85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is hybrid search?

Hybrid search is a retrieval method that runs keyword search, usually BM25, and vector similarity search over the same corpus and merges the two ranked lists, so a query matches both exact terms and meaning.

Checked

Vector search finds passages that mean the same thing as the query even when the words differ. Keyword search finds passages that contain the exact words, which matters for part numbers, error codes, names and clause references. Each one misses what the other catches.

Hybrid search runs both and merges the results. In 2026 it stopped being a specialist feature: vector databases added BM25, and search engines that started with BM25 added vectors.

Why one method is not enough

  • Where vectors miss

    Exact identifiers such as an SKU, a ticket number or a statute section. An embedding places them near similar-looking strings rather than matching them exactly.

  • Where keywords miss

    Paraphrase. A question about "cancelling a plan" will not match a passage about "ending a subscription" on keywords alone.

  • What the numbers show

    In Anthropic's Contextual Retrieval study, contextual embeddings alone cut top-20 retrieval failures by 35%. Adding contextual BM25 on top raised the cut to 49%.

How the two result lists are merged

  • Reciprocal rank fusion

    Each result scores by its rank in each list rather than its raw score, which avoids comparing BM25 scores with similarity scores on different scales.

  • Weighted blending

    Normalised scores from both searches are combined with a weight that is tuned on real queries for the corpus.

  • Reranking on top

    The merged shortlist goes to a reranker that reads the query and each passage together. In Anthropic's study this step took the reduction in failures to 67%.

Where hybrid search ships now

Pinecone made full-text search generally available on 2026-09-09, Qdrant 1.20 added BM25 scoring on full-text indexes on 2026-10-22, and turbopuffer shipped its second full-text engine on 2025-12-04. In Postgres, Tiger Data's pg_textsearch 1.0 (2026-03-31) brings BM25 next to pgvector. Elasticsearch, OpenSearch and Vespa came from the keyword side and have run both for years.

The practical result is that most teams no longer need two systems and a sync job to get hybrid search. The choice is which store already holds the data.

How Motiontic's engineers apply it

Motiontic's engineers build an eval set from the client's real questions before choosing weights or a fusion method, and include the awkward ones: exact codes, acronyms and questions phrased nothing like the source. Keyword-only, vector-only and hybrid runs are scored on that set, and the configuration that wins on the client's data is the one that ships.

Keyword search, vector search and hybrid search compared

Keyword search, vector search and hybrid search compared
Keyword (BM25)VectorHybrid
Matches onExact termsMeaningBoth
Strong atIDs, codes, names, rare termsParaphrase and conceptual questionsMixed real-world queries
Weak atSynonyms and rephrasingExact identifiersNeeds fusion tuning
NeedsAn inverted indexAn embedding model and vector indexBoth, plus a merge step

Common questions

What is the difference between hybrid search and vector search?

Vector search ranks passages by how close their embeddings are to the query's embedding. Hybrid search also runs keyword search, usually BM25, and merges the two lists, so exact terms such as product codes are not lost.

What is reciprocal rank fusion?

Reciprocal rank fusion merges ranked lists by giving each result a score based on its position in each list. It is a common way to combine keyword and vector results without calibrating their raw scores against each other.

Does hybrid search improve RAG?

In Anthropic's Contextual Retrieval study it did: adding contextual BM25 to contextual embeddings cut top-20 retrieval failures by 49%, compared with 35% for contextual embeddings alone.

Does hybrid search need a separate search engine?

Usually not in 2026. Pinecone, Qdrant, turbopuffer and Postgres (through pg_textsearch) all added BM25 alongside vectors, and Elasticsearch and OpenSearch support both.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.