Semantic Search
Search by meaning instead of matching words.
Retrieval that ranks results by embedding similarity rather than lexical overlap, so it finds relevant items that use different vocabulary than the query.
Semantic search is strong on paraphrase, weak on exact identifiers — SKUs, error codes, names. That is exactly why hybrid search exists.
Its characteristic failure is the plausible neighbour: a result that is semantically adjacent but factually wrong for the question. This is why production systems pair it with lexical search and reranking, and why evaluation of retrieval quality has to look at what was returned, not just at whether the model produced a fluent answer.
The near miss that reads well
A policy question returns the wrong policy’s section, which sounds similar.
Vectors captured the topic but not the distinguishing detail — a date, a plan name, a version. Lexical search catches exactly those tokens, which is the concrete argument for hybrid retrieval rather than a purely semantic index.
- docsGet text embeddings ↗
Embeddings over text with semantic search, classification and clustering as the canonical uses — plus the vector dimensionality that decides your index cost.
- paperColBERT: efficient and effective passage search via contextualized late interaction over BERT ↗
Late interaction: encode query and documents separately, then compare token-level representations at retrieval time. The accuracy/latency middle ground in neural ranking.
Embedding · Vector
A list of numbers representing meaning, so similar things land close together.
Vector Database · Vector store
Storage built for finding the nearest vectors, fast.
Hybrid Search
Keyword search and vector search, combined and merged.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.