Reproduce “Tool-to-Agent Retrieval: Bridging Tools and Agents for Scalable LLM Multi-Agent Systems” on OpenResearch

About this paper

Elias Lumer, Faheem Nizar, Anmol Gulati, Pradeep Honaganahalli Basavaraju, Vamse Kumar Subbiah

A research team at PwC U.S.A. developed a "Tool-to-Agent Retrieval" framework that effectively routes user queries to the most relevant tools or multi-agent systems by embedding both into a unified vector space. This approach achieved a Recall@5 of 0.83, an 18.6% relative improvement over prior methods, and demonstrated robust performance across diverse embedding models for large-scale LLM multi-agent systems.

Key insights

• Integrating tool-level granularity with agent-level contextual coherence in a unified index mitigates the problem of 'context dilution' and significantly improves selection accuracy.
• The ability to retrieve either a specific tool or an entire agent bundle in a single pass enables more adaptive and efficient routing decisions for LLM agents, avoiding brittle multi-stage pipelines.
• The architectural design of the unified indexing and retrieval mechanism provides robust performance gains across various proprietary and open-source embedding models, demonstrating its broad applicability.

Summary by alphaXiv.