Mistral Embeddings And Retrieval

Team builds a retrieval index with mistral-embed but later embeds new queries with a different embedding model, then compares vectors.

All vectors in an index must come from the same embedding model; query vectors must be produced by mistral-embed if the index was built with mistral-embed. Mixing models makes similarity meaningless.

Engineer hard-codes a vector dimension for mistral-embed copied from a forum post and provisions the vector store to that size.

Read the embedding dimensionality from an actual mistral-embed response and provision the store to match; do not hard-code a number from an unofficial source. Mark any cited dimension [REQUIRES-VERIFICATION] until confirmed against a live response.

Retrieval ranks by raw dot product over embeddings without confirming normalization and gets inconsistent ordering versus cosine.

Confirm whether mistral-embed vectors are unit-normalized; if not guaranteed, normalize consistently before using cosine, and ensure the vector store's distance metric matches. Treat the normalization assumption as [REQUIRES-VERIFICATION].