2 / 6 RAG Architecture & Retrieval

SKU Goes Missing

Support agents search your knowledge base by pasting exact tokens such as an error code (ERR_4021) or a part number. Pure dense vector retrieval returns loosely related troubleshooting prose and misses the one page containing the literal string. What is the standard fix?

A

Fine-tune the embedding model on past agent queries so the vector space memorizes each identifier and lexical matching stops being necessary at all

B

Run a lexical BM25 query alongside the vector query and fuse the two ranked lists with reciprocal rank fusion before reranking

C

Raise the ANN search effort parameter so the index explores more graph neighbours per query and recall over the vector space climbs

D

Drop the similarity score threshold so more of the vector neighbourhood survives into the candidate list handed to the reranker