Keyword, vector & hybrid search
Which passages should enter the candidate set?
Retrieval selects candidates from an index. Keyword search uses lexical matches; vector search uses similarity between representations. Hybrid search combines their results, often by fusing ranked lists. Filters constrain eligible records, while the candidate limit controls how much material later stages can inspect.
Query“laptop battery”
Laptop battery replacementExact words and related meaningFound
Notebook power-cell repairRelated meaning, different wordsSkipped
Laptop sleeve buying guideShares a word, misses the intentFound
Word matching can retrieve a result that shares vocabulary but misses the question.
Illustrative candidate sets. Real retrieval is model- and query-dependent.
What to understand
- Exact identifiers, names, and error codes can benefit from lexical matching; paraphrases may benefit from semantic matching.
- Hybrid search combines complementary signals. Rank fusion avoids assuming that lexical and vector scores share the same scale.
- Measure what the candidate set misses. A later reranker cannot recover a passage that retrieval never supplied.
Go to the source
Primary documentation for the ideas in this explainer.
Follow the next part of the system.
Reranking