Title: SkillPager: Enhancing Intra-Skill Navigation Through Query-Adaptive Semantic Node Retrieval

Abstract:

As Large Language Model (LLM) agents increasingly depend on extensive procedural documentation, a significant challenge arises: utilizing entire documents for prompting is inefficient, consuming excessive tokens while obscuring the specific information necessary for task execution. To address this, we define the problem of "intra-skill retrieval," which aims to extract a concise, execution-ready context from a known skill document based on a specific query.

We introduce SkillPager, a novel two-stage framework designed to optimize this process. The system operates by first parsing Markdown-based skills into typed semantic nodes during an offline phase. Subsequently, it employs Maximal Marginal Relevance (MMR) to conduct global, query-conditioned node selection in real-time.

Our evaluation on a benchmark comprising 1,975 queries across 395 skills demonstrates the framework's efficacy. SkillPager achieved a context sufficiency rate of 78.89% as judged by LLMs, falling slightly short of the 82.23% sufficiency rate attained by the exhaustive full-document baseline. Crucially, this performance was achieved with a 47.04% reduction in prompt token usage.

To isolate the drivers of efficiency, we conducted a granularity ablation study. We found that applying the same retrieval algorithm to raw, fixed-length chunks yielded a comparable sufficiency score of 81.77%. However, this approach resulted in a 28.81% increase in token costs compared to SkillPager. This discrepancy confirms that the efficiency improvements stem primarily from the use of typed semantic granularity rather than the retrieval algorithm itself.

When compared to graph-based alternatives, SkillPager surpassed the strongest baseline by a margin of 12.16%. Additional ablation studies revealed that supporting content is most effective when it is retained within the candidate pool for adaptive selection, rather than being discarded by static heuristics. Collectively, these findings highlight typed intra-document retrieval as a unique and critical access problem for skill-based agents.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC