Glossary
What is hybrid search?
In short
Hybrid search is the combination of search by words and search by meaning, with the results merged into one list.
Search and RAG · hybrid search
MMOOZZGG.com editorsChecked: 1 min read
For a child
You are looking for your mug in the school canteen. You can look for the label with your name, or by how it looks: blue, with a cat. If you search both ways at once, you find it faster, even if the label has worn off. Hybrid search does the same: it looks at both the words and the meaning, and then puts the finds together.
In plain words
Keyword search is good at exact matches: a part number, a contract number, a surname. Search by meaning understands rephrased questions but may miss a rare code. Hybrid search runs both methods on one query and merges the two lists into one. A document that ranks high in both lists rises to the top. For example, a manager types “invoice for Romashka LLC for VVG cable”, and the system finds the right client by name and similar invoices by meaning.
For business
Company documents hold many codes, part numbers and order numbers, and pure semantic search loses them. Combining the two methods closes this gap without large spending: you need a second keyword index and code that merges the lists. Many search engines and databases do this out of the box. The risk is in the weights: if you set wrongly which method matters more, the results get worse than either method alone. So the weights are tuned on a set of real queries where the right document is known in advance. For searching product catalogs and documents in 1C, Russia’s common accounting system, this approach is almost a must.
For a specialist
Hybrid search combines lexical search over sparse representations, usually BM25, with dense search over embeddings. The lexical part is precise on rare terms and identifiers, the dense part on paraphrases, so their errors overlap little. The lists are merged in two ways. The first is a weighted sum of normalized scores, where the weight of the dense part is tuned on validation. The second is Reciprocal Rank Fusion (Cormack, Clarke, Büttcher, SIGIR 2009), where a document gets the sum of 1/(k + r) over all lists, r is its position, and the authors fixed k = 60 in a pilot study. RRF needs no calibration of scores from different systems, so it is a convenient default. In Anthropic’s measurements (2024), adding contextual BM25 to contextual embeddings raised the reduction in failed top-20 retrievals from 35 to 49%. The merged results are then usually passed to a reranker.
Example
An engineer enters a bearing number from an old specification and adds “replacement with an equivalent”. Keyword search finds the card with this number, search by meaning adds supplier letters about equivalents, and both results land in one list.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map