PA-RAG: Protocol-Aware Retrieval-Augmented Generation for Tool Discovery in MCP Ecosystems

ACL ARR 2026 January Submission3351 Authors

Keywords: Retrieval-augmented generation, Passage retrieval, LLM/AI agents, Model Context Protocol

Abstract: MCP has emerged as foundational infrastructure for AI-native applications since its inception. The ecosystem's rapid expansion makes it a significant challenge to discover tools across thousands of servers precisely. To address it, we propose Protocol-Aware Retrieval-Augmented Generation (PA-RAG) for Tool Discovery in MCP Ecosystems. PA-RAG utilizes topology-aware hierarchical indexing to model structural relationships between MCP servers and tools. This indexing facilitates multi-stage cascade retrieval across hierarchical layers. PA-RAG adopts a complexity-adaptive dual-path execution mechanism to handle varying query types. Simple queries employ direct RAG. Complex queries follow a decomposition-aggregation strategy to maximize functional tool coverage. Extensive experiments based on lightweight LLMs demonstrate that PA-RAG significantly outperforms the state-of-the-art, effectively balancing discovery latency, coverage, and precision.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: Retrieval-augmented generation, Passage retrieval, LLM/AI agents

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 3351