Semantic Search — AI Glossary

Reference details

Related terms

• Embedding
• Vector Database
• Hybrid Search
• BM25

How it works

1. Documents are split into chunks and each chunk is passed through an encoder model, producing a fixed-length vector whose direction encodes meaning rather than wording.
2. Those vectors are written to an index built for approximate nearest-neighbour search, because scanning every vector per query does not scale past small corpora.
3. At query time the question is embedded by the same encoder — a different one places its output in an unrelated coordinate space and the comparison becomes meaningless.
4. Similarity is computed as a dot product or cosine between the query vector and candidate vectors, which is why vectors are usually normalised at insertion so the two agree.
5. The index returns the top-k nearest chunks, and the ranking is by geometric proximity alone: nothing in the mechanism knows what the words were.
6. That last property is the whole character of the method — matching is a distance calculation in a learned space, so anything the encoder did not place in that space cannot be retrieved by it.

Example: The query it gets wrong, and what share of traffic that is

“Documents that do not mention pricing” — semantic search returns documents about pricing, because negation barely registers in embedding space and the topic dominates. The same weakness explains why “increase” and “decrease” retrieve each other: they are about the same thing while meaning opposites. The question that matters operationally is how much traffic that costs. Sample the query log: if 8% of queries contain a negation or an explicit opposition, and the retriever returns topically-right, semantically-inverted results on 70% of those, then 0.08 x 0.70 = 5.6% of all queries return confidently wrong context. That is not a small defect rate for a retrieval layer, and it is invisible in aggregate relevance scores because the returned documents are about the right subject. It also points at the fix: 5.6% is a routing problem, not an embedding problem — detect the negation and send those queries to a lexical path with a filter.

Commonly confused with

Semantic search is the goal, not the technique. Dense bi-encoder retrieval is the usual implementation, but sparse methods like SPLADE also target meaning while staying sparse. Conflating the goal with one implementation hides that hybrid approaches exist. The distinction has a practical edge: stated as a technique, the response to a recall failure is to change embedding model, when the better response is often to add a second retriever whose failure modes differ.

When to use it

Reach for it when:

• Users describe what they want rather than naming it
• Vocabulary differs between query and document
• Concept-level recall matters more than exact matching
• Alongside a lexical retriever, so the two failure modes do not coincide

Reach for something else when:

• The query is an identifier, code or exact name
• Negation or opposition must be distinguished
• You need an auditable reason why a result was returned
• Without measuring the negation share of your query log, which was 5.6% of traffic returning wrong context above

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Semantic Search — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path