Semantic Search
Semantic search is search that matches a query to text by meaning rather than by shared words, so a relevant passage can be returned even when it has no vocabulary in common with the query.
The technique is defined by what it fixes. Keyword search finds documents containing the query’s terms; if the document says “invoice remittance” and the query says “how do I pay this bill,” the words never meet and the document is not found. Semantic search closes that gap by comparing representations of meaning instead of surface strings.
In current systems the representation is almost always an embedding. Query and text are each converted into a vector by the same model, and proximity in that space stands in for similarity of meaning. This is why the phrase vector search is so often used as a synonym, though it names the mechanism rather than the goal.
The term is older than embedding-based retrieval and has covered several mechanisms over time, including ontologies, thesauri, and query expansion by synonym lists. What has stayed constant is the objective: retrieval that survives paraphrase.
In practice
A semantic search over a text collection has a fixed shape. The collection is chunked, each chunk is embedded, and the vectors go into an index. At query time the query is embedded by the same model and the nearest stored vectors are returned, usually with a limit on how many.
Its characteristic strengths and failures both follow directly from matching on meaning rather than on words.
It handles paraphrase, synonymy, and questions phrased unlike the source. It tolerates some misspelling, because a misspelled word often still embeds near its correct form. It works across languages when the embedding model was trained to place translations near each other.
It handles exact identifiers badly. A part number, an error code, a person’s surname, or a version string carries meaning that is precisely not semantic — the value is the exact string, and an embedding model has no reason to distinguish one similar-looking code from another. Queries of that kind routinely return plausible neighbours instead of the exact match.
It also always returns something. Nearest-neighbour search over a non-empty index yields results regardless of whether any of them are relevant, so an off-topic query produces the closest available chunks rather than nothing. Distinguishing “these are the best matches” from “these are good matches” requires a separate judgement.
Commonly confused with
Vector search. The mechanism, not the goal: searching by vector proximity. Almost all semantic search is implemented as vector search, but vector search over non-semantic vectors — image features, audio fingerprints, learned user representations — is not semantic search. The two words are used interchangeably in practice and the substitution is usually harmless.
RAG. Retrieval-augmented generation uses retrieval to supply a language model with material, then generates an answer. Semantic search is one way to do the retrieval part. A semantic search product returns a ranked list of passages; a RAG system returns prose. Treating them as the same thing is the most common category error in the area.
Hybrid search. Semantic and keyword retrieval run together and their results merged. It contains semantic search rather than competing with it. See hybrid search.
Dense retrieval. The technical name for the same operation in information-retrieval writing, contrasted with sparse retrieval. “Semantic search” is the product-facing term; “dense retrieval” is the literature-facing one. See dense retrieval.
Natural language search. A claim about the input — that a user may type a sentence rather than keywords. That is compatible with several retrieval mechanisms, including ones that parse the sentence into a structured query, so it does not imply semantic matching.
Usage notes
The phrase is heavily used in marketing and frequently means only that a product embeds text. It carries no commitment to a particular quality level, and a system described as semantic can be worse on a given corpus than well-configured keyword search — most visibly on collections dominated by identifiers and jargon.
“Semantic” here does not mean the Semantic Web. That earlier usage referred to explicit machine-readable structure, ontologies, and linked data — a different technical tradition that also claimed the word. Documentation from that lineage uses “semantic search” to mean querying a knowledge graph.
Relevance is not symmetric with similarity. Embedding proximity measures resemblance, and a query resembles a document that quotes it more than one that answers it. A question and its answer are not paraphrases of each other, which is why some models are trained specifically for question-to-passage matching rather than general similarity.
Threshold values do not transfer. A similarity cut-off tuned on one model and one corpus is not meaningful on another. See cosine similarity.