Embed queries and documents with separate optimizations — asymmetric embeddings beat one-size-fits-all
StackAI·medium signal
Queries are short and documents are long, so providers that optimize the two embedding paths separately consistently improve retrieval over symmetric embedding of both. Models like BGE-M3 are favored for production RAG because they handle this asymmetry and hold up across diverse corpora. If your retriever embeds queries and passages identically, switching to an asymmetric/query-document-aware model is a near-free recall gain.