What does an embedding capture?

An embedding model maps content to a vector of numbers. Items with related patterns may be close under a chosen similarity measure. The representation depends on the model and training data; closeness does not guarantee factual equivalence.

How is it used in retrieval?

A system embeds document chunks, stores their vectors and source metadata, and embeds a new query with a compatible model. A vector index returns nearby candidates, which may be filtered or reranked before an answer is generated. Re-embedding may be needed when models change. Vector indexing for LLM applications explains this use in search.

Why can similar vectors disagree on facts?

Two passages about the same product may state opposite prices or dates and still be close in vector space. Source verification and precise comparison remain necessary. For long documents, chunking usually retains local topics better than one embedding for the whole file.

FAQ

Is an embedding a vector database?

No. The embedding is the representation; a database stores and searches it.

Does high similarity mean factual agreement?

No. Related passages may contradict each other.

Can a whole document use one vector?

Sometimes, but long documents usually need chunk level representations.