Semantic Search: Finding Content by Meaning Instead of Keywords
Semantic search uses vector embeddings to match content by meaning rather than exact keyword overlap, enabling retrieval of conceptually related documents with zero shared words.
Semantic search is a retrieval approach that matches documents by meaning rather than exact keyword overlap. Text is encoded into dense vector embeddings — numerical representations in high-dimensional space — using transformer models. Queries are embedded the same way, and nearest neighbors are found by cosine similarity or dot-product distance. This enables finding conceptually related content even with zero shared words. A query for "how to fix a broken computer screen" can match a document about "laptop display replacement procedures" — something keyword search misses entirely. Modern production systems typically use hybrid search: combining semantic (embedding) search with traditional BM25/TF-IDF keyword ranking, since each covers different failure modes. Keyword search excels at exact terms, names, and codes; semantic search handles paraphrases, synonyms, and conceptual queries. Key infrastructure includes vector databases (pgvector: PostgreSQL Extension for Vector Similarity Search, Pinecone, Weaviate, Qdrant) that index embeddings for fast approximate nearest-neighbor search, and embedding models (OpenAI, Cohere, open-source alternatives) that convert text to vectors. **See also:** The LLM Wiki Pattern: Personal Knowledge Bases Without RAG · Vector Databases: How Embedding Search Powers Modern AI Applications