Glossary

Hybrid search

Hybrid search is a retrieval method that runs keyword (lexical) search and vector (semantic) search over the same query at the same time, then merges the two ranked result lists into one, instead of relying on either alone. Keyword search keeps exact-match precision — a SKU number, brand name, or model number has to match character for character to be useful. Vector search adds meaning-based recall — a query like "bolt that won't rust outdoors" can surface a 316 stainless hex cap screw even though no word overlaps. Run alone, each method has a blind spot the other covers, which is why most product search engines built in the last few years default to running both.

Why neither method alone is enough for a product catalog

Keyword, or lexical, search — commonly scored with an algorithm like BM25 — is precise and exact-match: excellent for SKU numbers, model numbers, and brand names, but it returns nothing for a query that shares no words with the product record it should match, like "something to stop a door from rusting" against a listing titled "316 Stainless Hex Cap Screw."

Vector, or semantic, search converts the query and every product record into embeddings and matches by meaning rather than text overlap, which catches paraphrasing and intent. Its blind spot runs the other way: a vector model can rate two different model numbers as "similar" because the surrounding description text is similar, which is exactly the case where a buyer wants zero tolerance for substitution.

How the two result lists get merged

The most common fusion method is Reciprocal Rank Fusion (RRF), which merges on rank position rather than raw score. Each retriever (keyword and vector) returns its own ranked list; a document's fused score is the sum of 1/(k + rank) across every list it appears in, where k is a constant, commonly set to 60, and a document absent from one list simply contributes nothing from that side.

Ranks are used instead of raw scores because the two methods don't produce comparable numbers: BM25 scores are unbounded and depend on query length and term statistics, while cosine similarity over normalized embeddings falls in a fixed range. Comparing them directly means comparing numbers on different scales. RRF traces back to a 2009 SIGIR paper by Cormack, Clarke, and Buettcher, and it's now the default hybrid-fusion method in several widely used search and vector platforms, including OpenSearch, Elasticsearch, and Azure AI Search. A weighted linear combination of normalized scores is the main alternative where a team wants more control over how much weight each side gets.

What it costs, and where it depends on clean product data

Hybrid search means maintaining two indexes against the same product records — a lexical index and a vector index — and keeping both current as the catalog changes, which adds infrastructure and query latency over running either method alone.

It's also only as good as the product data feeding both sides. A vector index built on thin, inconsistent, or boilerplate descriptions embeds that noise into the semantic layer. A keyword index built on inconsistently normalized attribute values — "Zinc Plated" in one record, "Zinc-Plated" in another — returns inconsistent exact matches regardless of which fusion method sits on top. Neither retrieval method, nor the fusion between them, can compensate for catalog content a PIM never normalized in the first place.

Frequently asked questions

What is hybrid search?

Hybrid search runs keyword (lexical) and vector (semantic) search over the same query at the same time and merges their ranked results into one list. It combines keyword search's exact-match precision with vector search's ability to match on meaning rather than shared words.

How does hybrid search combine keyword and vector results?

The most common approach, Reciprocal Rank Fusion, scores each document by its rank position in each result list rather than by raw score, since keyword and vector scores aren't on comparable scales. A document ranked highly by both methods typically beats a document ranked #1 by only one of them.

What is Reciprocal Rank Fusion?

Reciprocal Rank Fusion (RRF) is a method for merging multiple ranked result lists by summing 1/(k + rank) for each document across every list it appears in, with k typically set to 60. It originates from a 2009 SIGIR paper and is the default hybrid-search fusion method in platforms including OpenSearch, Elasticsearch, and Azure AI Search.

Does hybrid search need good product data to work well?

Yes. Hybrid search runs two indexes — lexical and vector — over the same underlying product records, and both inherit any weaknesses in that data. Thin descriptions produce noisy embeddings, and inconsistently normalized attribute values produce inconsistent keyword matches, regardless of how well the fusion layer merges the two result sets.

Related terms

See it on your own SKUs.

A 30-minute walkthrough on your categories and your supplier data.

Book a demo