Semantic Search
Finding information by meaning, not by keyword — the foundation of modern enterprise search and RAG.
Semantic search is a retrieval technique that finds content based on meaning rather than exact keyword matches. It works by converting both the query and the documents into vector embeddings — high-dimensional numerical representations of semantic meaning — and finding the documents whose embeddings are closest to the query embedding.
Keyword vs Semantic
Keyword search matches exact words: a search for 'timber construction' finds documents containing those exact words. Semantic search matches meaning: the same query also finds documents about 'Holzbau' (German for timber construction) or 'wooden building methods' because their embeddings are semantically similar. Neither approach is strictly better — they solve different problems. Keyword search is precise; semantic search is flexible.
How It Works
- The query text is passed through an embedding model, which produces a vector (typically 768-3072 dimensions).
- The vector is compared against all document vectors in the database using cosine similarity or dot product.
- The top-N most similar documents are returned as candidates.
- Candidates are re-ranked by a cross-encoder for true relevance (similarity is not the same as relevance).
Semantic search is not magic — it depends on the quality of the embedding model and the relevance of the indexed content. A good embedding model captures domain-specific meaning; a generic one may miss nuance. The Knowledge Hub allows per-workspace embedding model selection for this reason.
FAQ
Frequently Asked Questions
The Knowledge Hub supports multiple embedding models (e.g., text-embedding-3-large, Cohere embed). The model is configurable per workspace. Domain-specific models can be integrated for specialised use cases (medical, legal, technical).
Retrieval from a vector database with approximate nearest neighbour (ANN) search typically takes 10-50 milliseconds for corpora up to millions of documents. Re-ranking adds 50-200ms depending on the number of candidates. Total retrieval latency is well under 500ms in most cases.
Related Topics
Fundamentals
What is RAG?
Retrieval-Augmented Generation — the technique that grounds LLM responses in verified, attributable sources.
Fundamentals
Vector Embeddings
High-dimensional numerical representations of text that capture semantic meaning for machine retrieval.
Implementation
Hybrid Search
Combining semantic and keyword search to get the best of both — meaning and precision.