Embeddings

An embedding model is trained so that semantically related inputs receive nearby vectors, typically measured by cosine similarity. Because the comparison is geometric rather than lexical, "how do I cancel my plan" can retrieve a passage titled "ending a subscription" with no shared keywords. The vectors themselves are opaque: no individual dimension means anything you can name, and you cannot reconstruct the original text from them, though you can often infer a surprising amount about it.

Two operational details decide whether an embedding pipeline works. First, the same model must be used for indexing and for querying — vectors from different models occupy incompatible spaces, and mixing them produces confidently wrong neighbours rather than an error. Second, what you embed matters more than which model you pick. A chunk that spans two unrelated topics gets an averaged vector that is close to neither, which is why chunking strategy is usually the highest-leverage knob in a retrieval system.

The misconception to avoid is treating similarity as relevance. Embeddings capture topical closeness, not correctness, recency or authority. A three-year-old deprecated policy and this morning’s replacement will sit almost on top of each other in vector space. Metadata filtering, recency weighting and reranking exist precisely because distance alone is not enough to decide what belongs in the answer.