How do vectors represent related content?
An embedding model maps text or other content into a numeric space. Nearby vectors may capture paraphrases or related concepts. Similarity is statistical and does not prove agreement about dates, amounts, or conclusions.
What happens during indexing and search?
The system creates embeddings for items or chunks, stores identifiers and metadata, and builds a nearest neighbor index. It encodes the query, searches, filters, and may rerank candidates. Approximate indexes trade some recall for speed, so settings need testing. See vector indexing for LLM applications.
Which queries still need exact terms?
Product IDs, policy numbers, error codes, and precise names often require lexical matching. Keyword search and metadata filters preserve those constraints. Vector search is usually one part of a retrieval workflow, and hybrid search can combine it with exact matching.
FAQ
Is a dedicated vector database required?
No. Relational, search, and analytical databases may support vector indexes.
Are more vectors always better?
No. Chunking, model choice, index settings, and cost matter.
How is a similarity threshold chosen?
Measure false matches and misses on the target collection.