Vector search

First, every passage is turned into an embedding and stored in a vector database. When a question comes in, it gets an embedding too.

The database then returns the stored passages that sit closest to the question.

A cooking site stores each recipe step as a separate passage. Someone asks: "How do I stop pasta from clumping?" Vector search can return the step about stirring and using plenty of water, even though that step never uses the word "clumping".

Anthropic describes this as the usual setup for RAG: split documents into chunks, turn them into embeddings, store them in a vector database and search by meaning.

Anthropic also notes that embeddings can miss exact matches, such as an error code. That is why many systems add keyword search too.

• Write passages that make sense on their own, so a single retrieved piece still answers the question.
• Write exact names, model numbers and codes in plain text, because meaning-based search alone can miss them.
• Use clear headings, so each passage has a clear topic.

Semantic search

Semantic search is the goal: matching meaning. Vector search is one common way to get there.

Hybrid search

Hybrid search combines vector search with keyword search.