SCIENCE · RETRIEVAL

What is hybrid search?

AUTHOR · XERJ documentation team · REVIEWED BY · XERJ engineering team · PUBLISHED · 2026-11-13 · UPDATED · 2026-11-14

SHORT ANSWER · UPDATED 2026-11-14

What is hybrid search? It runs a full-text query, such as BM25, and a vector query on the same question, then merges the two ranked lists. Full-text search catches exact names and error strings. A vector query can catch the same idea in different words. XERJ can fuse the two arms with reciprocal rank fusion.

TL;DR Hybrid search runs two searches on the same query: a full-text search such as BM25, and a vector search. Each returns a ranked list. A fusion step, often reciprocal rank fusion, merges them into one list.

Try it: two lists, one ranking

How hybrid search puts lists together

Match the words

BM25 scores a document by query terms, their frequency, and how rare they are. It is strong when the words in the query appear in the passage.

The fusion method is reciprocal rank fusion, as described by Cormack, Clarke and Buettcher (SIGIR 2009). How XERJ turns this on is on the hybrid-search answer. This figure does not report a new score.

PASTE THIS TO YOUR AI AGENT — IT DOES THE REST

Act as a coding agent. Read https://xerj.org/llms.txt, install XERJ, start a local node, index one folder I point you to, and run one search so I can see a cited passage.

How it works

Full-text search scores a document by the query terms it contains, how often they appear, and how rare they are. It is strong when the words in the query appear in the passage. It misses a passage that says the same thing in different words.

A vector query can rank that paraphrase, because closeness in the vector space does not require shared terms. It is weaker on exact tokens, where two similar identifiers can sit close together.

Hybrid search runs both, then fuses the ranks. Reciprocal rank fusion is a common way to do that without mixing raw scores that sit on different scales. The worked example is on the reciprocal rank fusion page.

How XERJ uses it

XERJ keeps a lexical index and vectors in the same engine. The default embedder is lexical and runs offline. Neural mode adds a trained vector arm. You can fuse a lexical arm with a vector arm in one request.

The how-to, including the request shape and when to use RRF or a weighted sum, is how to run hybrid search in XERJ. A second practical page is how XERJ combines search.

See also BM25, semantic search, and reranking.

History and research

The lexical arm in most hybrid systems is BM25, surveyed by Robertson and Zaragoza in The Probabilistic Relevance Framework: BM25 and Beyond (2009).

Dense vector retrieval for question answering is described by Karpukhin and colleagues in Dense Passage Retrieval for Open-Domain Question Answering (EMNLP 2020).

The usual fusion method is reciprocal rank fusion, introduced by Cormack, Clarke and Buettcher in Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods (SIGIR 2009). This page does not treat those papers' scores as XERJ measurements.

Guides

Run it yourself with the hybrid search recipe. The hybrid search answer shows one request, and how XERJ combines search explains the merge.

FAQ