Section 2

Retrieval Quality

A retriever that is good on average can fail systematically on the questions a team cares most about, such as error codes pasted by service-desk staff. This section measures retrieval per slice on a labelled set drawn from the knowledge base, fuses BM25 with dense retrieval, prices a reranker at 12,000 questions a day and compiles the access filters that decide which chunks a question may see, showing where each stage helps and where it hurts.