What is semantic search?

Semantic search matches queries to documents by meaning rather than by exact keyword overlap. Instead of scoring pages based on whether they contain the literal search terms, the system encodes both the query and candidate documents as dense vectors using an embedding model, then ranks results by vector similarity. A query for "cheap cloud storage" returns results about "affordable object storage" even if neither phrase appears verbatim in the other, because the vectors for both are close in the embedding space. Traditional keyword-based search ranks by term frequency; semantic search ranks by conceptual proximity.

Use semantic search when queries are in natural language, when users phrase questions differently from how documents are written, or when building a retrieval pipeline for a RAG grounding system where stored document terminology may not match incoming queries exactly. Keyword search remains the better choice when precision matters more than recall, such as searching for a specific product code, API parameter name, or error message string.

Firecrawl's Search API surfaces relevant pages across the indexed web. Pair it with an embedding model and a vector store to build a semantic retrieval pipeline: Firecrawl finds candidate pages and extracts clean Markdown, and the embedding layer handles similarity ranking over the content.