Embeddings are what let a RAG pipeline find text by meaning instead of exact words. Each chunk becomes a vector — a list of numbers — and passages with similar meaning land near each other in that space.
Why meaning beats keywords
A keyword index can't connect "how do I sign in?" to a doc titled "authentication." Embeddings can, because both map to nearby vectors. That's the whole point of semantic search — the user's words don't have to match the document's words.
Choosing a model and dimension
Smaller models (384–768 dimensions) are fast, cheap, and good enough for most retrieval; larger ones cost more memory and compute for marginal gains. Pick one and keep it consistent — you can't mix embeddings from different models in the same index, since their vector spaces don't line up.
Embeddings are necessary, not sufficient
Vector similarity alone still misses exact terms like error codes or product names, which is why production systems pair it with keyword matching in hybrid search — and why good RAG chunking matters, since you're embedding chunks, not whole documents. Kit for AI embeds every chunk on ingest and searches it for you. Start free and try a real query.