Turning a corpus into a skill tree cuts RAG hallucination from 50% to 4.5%

Notes on Don't Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG (arXiv:2604.14572) — Yiqun Sun, Peng Wei, Lawrence B. Hsieh · April 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

Corpus2Skill turns a document corpus into something an agent walks instead of queries. An offline compile pass embeds every document, clusters them hierarchically with K-Means, and has an LLM write a summary at each level, producing a tree of SKILL.md and INDEX.md files plus a cross-branch entity index — the same filesystem format and progressive disclosure mechanism used for procedural agent skills, repurposed here to describe what a corpus contains rather than how to do a task. At serve time the agent is preloaded with only skill names and one-line descriptions, then drills from a top-level summary through progressively finer INDEX.md files down to a specific document, backing out of branches that turn out thin. Five baselines take the comparison on WixQA, an enterprise customer-support benchmark of 6,221 support articles and 200 expert-written queries with gold answers: BM25, dense embedding retrieval, hybrid retrieval, RAPTOR-style hierarchical retrieval, and an agentic RAG agent with search tools.

Navigation wins on every quality metric on WixQA. Token F1 lands at 0.456, a 21% relative gain over agentic RAG (0.378) and 25% over dense retrieval (0.364), with factuality 6 points above the next-best method. It grounds better too: hallucination rate falls to 4.5%, against 50% for the same agent searching instead of navigating — though it costs more, at $0.153 per query versus $0.082 for agentic RAG and $0.007–$0.012 for the single-shot retrievers. Swapping the serving model from Sonnet 4.6 to Haiku 4.5 cuts that to $0.093 (−39%) while keeping 92% of the F1 score, which suggests the compiled tree does most of the work, not the model walking it. A follow-up sweep across ten RAGBench subsets complicates the picture: navigation wins on five single-domain sets (by 0.01–0.14 F1), ties on two heterogeneous ones, and loses on three — an open-domain factoid pool, a homogeneous-tabular corpus, and a long-extractive-document set — by 0.04–0.17 F1, because their top-level cluster summaries collapse into generic, indistinguishable labels.