Glossary
Embeddings
Embeddings are lists of numbers that a model produces to represent the meaning of text, images or other content, placing similar meanings close together even when the wording differs. Each embedding belongs to the model that created it, and embeddings from different models cannot be compared.
From text to a point in space
A keyword search for staff turnover misses a report that only says employee attrition. An embedding model converts each phrase into a list of numbers, typically several hundred to a few thousand long, and the two phrases land as nearby points. Documents are embedded once and stored in a vector database. A query is embedded the same way at search time, and the closest stored points come back. That is the mechanism behind semantic search and the retrieval step of retrieval-augmented generation. Some models place text and images in one shared space, so a written query can find a picture.
Why is an embedding tied to one model?
Each model builds its own space. The same sentence gets different numbers from different models, and distances only mean something inside one space. A corpus and the queries against it must therefore come from the same model and version, and tagging every stored vector with both is how mixed spaces get caught.
A model upgrade means re-embedding everything
Switching to a better embedding model invalidates every existing vector. The whole document collection is read back from storage and run through the new model, which for a large corpus can take days. Embeddings are a lossy summary and cannot be turned back into the original text, so the source documents have to be kept intact and readable for this to be possible at all.
What embeddings do not capture
They capture meaning, not truth or permission. Two contradictory statements on the same topic sit close together, and a restricted document embeds as easily as a public one. Exact strings blur, so part numbers and names that differ by one character land almost on top of each other, which is the gap hybrid search fills. Quality also depends on how text was split before embedding, through document chunking. Because a vector carries traces of its source text, embeddings of sensitive content need the same access control as the documents.