Standard RAG systems are bad at conversations — and a new paper puts a number on exactly how bad.
Researchers tested how well a typical retrieval-augmented generation setup handles enterprise support sessions, where users ask a series of related questions rather than one-off queries. A single retrieval call over an unmodified knowledge base covered only 41% of what a user actually needed across a full session. Their fix: reorganize the knowledge base offline using co-occurrence-aware clustering, then expand retrieval candidates through cluster neighborhoods at query time. On WixQA — a benchmark of 6,221 enterprise support articles — the approach pushed single-query session coverage to 58%, cut the number of retrieval calls needed to reach 70% coverage by 34%, and compressed the knowledge base to 20% of its original size.
The results hold across four embedding models and six functional domains, which matters: gains that only appear under one model or one topic area are hard to ship with confidence. More pointedly, the authors argue that session-level coverage should replace single-query recall as the primary benchmark for enterprise RAG — a framing shift that would make most current RAG deployments look worse on paper than vendors like to admit.
Most RAG benchmarks still reward systems for nailing isolated questions; this work suggests that benchmark flatters the wrong thing entirely.