Polp
← Back to glossary
RAG and search

Semantic search

Search that uses the meaning and context of a query to retrieve and rank relevant information, not only exact word matches.

Business example

An employee asks about holiday carry-over. Semantic search can retrieve the annual leave policy even when the document uses different wording, while permissions and metadata keep the result within the employee's allowed sources.

How it works

  1. 1.Understand the query's intent and represent its meaning with language models, embeddings, or both.
  2. 2.Retrieve an initial candidate set using semantic, keyword, or hybrid matching.
  3. 3.Apply metadata and permission filters before exposing results to the requester.
  4. 4.Optionally rewrite the query and rerank the strongest candidates for relevance.
  5. 5.Return ranked passages or documents that an answer layer can cite and evaluate separately.

Common misconceptions

Semantic search is always vector search with no keyword matching.
Production systems often combine semantic and lexical candidates, then filter and rerank them.
A high semantic relevance score proves the final AI answer is correct.
Ranking selects promising evidence; groundedness, completeness, citations, and answer quality still need separate evaluation.

Related concepts

EmbeddingsHybrid searchRerankingMetadata filtering

Related reading

Enterprise search in 2026: answers, permissions, evidence
How to evaluate internal AI search tools: a 15-question checklist

Sources

Reviewed:

Reviewed by: Javier Chulvi Bernad · LLM Engineer · Madrid