Una recepción no puede cubrir una sesión: el 41% de la cobertura revela el punto ciego de RAG

A single retrieval call over a standard knowledge base covers only 41% of a user's session-level information need. That means 59% of what a user actually wants across a coherent set of related questions is simply missing from that first pass. Enterprise RAG has been optimizing the wrong metric.

The 41% Gap That Enterprise RAG Ignores

Most RAG systems treat each query as an independent event. But enterprise users arrive with sessions—episodes of related questions that span semantically distant parts of the knowledge base. The authors tested this on WixQA, a corpus of 6,221 enterprise support articles, and found that standard single-query retrieval covers only 41% of what a user needs across a session. That’s not a recall problem; it’s a design problem.

Clustering documents by co-occurrence—training on which documents tend to be retrieved together across session histories—lets the system reorganize the knowledge base offline. At query time, retrieval expands to include cluster neighbors. The result: single-query session coverage jumps to 58%, a 17% absolute gain (95% CI: [14.1, 20.4]).

Clustering by Co-Occurrence: How It Works

Instead of reorganizing for each query, the system performs one offline pass: co-occurrence-aware clustering groups related documents based on session-level usage patterns rather than semantic similarity alone. Then, at retrieval time, the system expands candidates through cluster neighborhoods. This does something counterintuitive: reducing the number of retrieval calls needed to hit 70% session coverage by 34%, while compressing the effective KB to 20% of its original size.

That consistency holds across four embedding models and six functional domains. The method doesn’t depend on a particular encoder — the structure is what matters.

What This Means for Your Knowledge Base

If you’re building enterprise RAG, single-query recall is the wrong yardstick. Session coverage captures the reality of how users actually ask questions — in coherent chains, not isolated shots. The authors argue that session-level coverage should be the primary metric for enterprise RAG evaluation. This method gives you a concrete knob: reorganize the KB by co-occurrence clusters, expand retrieval to cluster neighborhoods, and measure coverage across sessions instead of individual queries.

Run the math on your own KB. If your users ask follow-up questions, 41% is where you start — and 58% is the minimum you should accept.

Source: One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG
Domain: arxiv.org