hard20 pts · first answer countsvector-searchrag

Filtered vector search returns too few

You request the top 10 nearest documents, then filter them to tenant = "acme". Users often get only 1 or 2 results even though acme has thousands of documents. Why?

A.The embedding model is too large for the indexB.Cosine similarity cannot work with metadataC.Filtering after top-k ANN drops most hits; use filter-aware searchD.The index needs more dimensions per vector

← Fine-tuned model forgets general skills Inputs stable but accuracy falls →