SkillPager: Query-Adaptive Intra-Skill Navigation via Semantic Node Retrieval

SkillPager is a novel retrieval framework that optimizes how large language model agents access long procedural documents by selecting minimal, execution-sufficient context from skill documents. The system achieves 78.89% sufficiency while reducing prompt tokens by 47.04% compared to full-document prompting, demonstrating that typed semantic granularity significantly improves efficiency in skill-based LLM agent systems.

SkillPager addresses a critical inefficiency in modern LLM agent architectures. As AI systems increasingly rely on procedural skill documents to execute complex tasks, the computational cost of full-document prompting creates substantial waste. This research tackles the problem through a two-stage approach: offline parsing of Markdown skills into typed semantic nodes, followed by online query-conditioned selection using Maximal Marginal Relevance algorithms. The methodology represents a meaningful advance in how AI systems consume contextual information.

The research emerges from growing pains in LLM agent development. As agents scale to handle more sophisticated tasks, the documents governing their behavior grow proportionally, making naive full-document inclusion economically and informationally suboptimal. SkillPager's 47% token reduction while maintaining near-baseline sufficiency levels directly addresses production cost concerns that plague deployed AI systems. The benchmark evaluation across 395 skills and 1,975 queries provides robust empirical grounding absent in much AI research.

For developers building LLM-based systems, this work has immediate practical implications. The gap between SkillPager's 78.89% sufficiency and the full-document baseline's 82.23% suggests a viable trade-off for cost-sensitive applications. The finding that typed semantic granularity outperforms fixed-length chunking indicates that document structure carries information value beyond raw text. This informs how organizations should architect their knowledge systems for AI agent consumption.

Future development likely focuses on extending these techniques to multi-document retrieval scenarios and exploring how semantic typing strategies generalize across different document formats and domains. The research establishes typed intra-document retrieval as a distinct optimization problem worthy of continued investigation.

• →SkillPager reduces prompt tokens by 47% while maintaining 78.89% LLM-judged sufficiency in skill document retrieval
• →Typed semantic node granularity outperforms fixed-length chunking, suggesting document structure carries inherent information value
• →Two-stage offline parsing and online MMR-based selection enables efficient query-adaptive context selection
• →Graph-based baselines underperform SkillPager by 12.16%, indicating superiority of the semantic retrieval approach
• →Supporting content retention and adaptive selection outperforms static heuristic filtering strategies