What is Vector Embedding?
- Vector Embedding
- A vector embedding is a list of numbers representing the meaning of a piece of text, positioned so that passages with similar meaning sit close together in the same space. Retrieval systems use vector embeddings to find relevant passages by proximity rather than by keyword match, which is how a page gets retrieved for a question that shares none of its words.
01 In practice
Keyword matching and embedding-based retrieval fail differently. A keyword system misses a passage that answers the question in other words; an embedding system finds that passage, and can also surface something merely adjacent in meaning, which is why precision in wording still matters.
Embeddings are computed over chunks rather than whole documents, so a page's retrievability is really the retrievability of its parts. A section covering three loosely related topics produces an embedding sitting between all three and close to none of them.
Writing implications follow from that geometry. One question per section, answered in the vocabulary a reader would actually use, produces a chunk with a clear position; hedged, multi-topic prose produces one that is never the nearest match to anything.
02 Related terms
The terms next to this one.
Content
Content Chunking
Structuring a page into sections that each stand alone.
Technical
Retrieval
A model fetching live pages while composing an answer.
Technical
Passage Ranking
Retrieval that scores sections of a page, not the whole page.
Core
Grounding
Tying a generated answer to sources it can be traced back to.
The work behind it
Tell the models what you are.
Markup that tells models what your business is, on pages that render without JavaScript.