FED-GRAPH-MCP: Scaling LLM Tool Discovery via Sparsified Functional Graphs and Incremental Synchronization

Assumes deep technical literacy. Bridges to the closest neighbouring fields.

**Problem:** As the Model Context Protocol (MCP) ecosystem scales past 10,000 servers, LLM agents face a severe tool-discovery bottleneck because existing dense retrieval methods treat tools as isolated documents and ignore functional interdependencies and real-time registry volatility. **Method:** FED-GRAPH-MCP addresses this via a hybrid retrieval architecture that constructs a multi-typed functional tool graph capturing relational dependencies, then fuses Relational GCN (R‑GCN) structural embeddings with dense semantic embeddings through a learned reranker. A CRUD-triggered incremental synchronization daemon enables efficient index updates without full recomputation. **Main Results:** On LiveMCPBench and MCP-Bench, graph sparsification reduces the in-memory structural footprint by 95.6%, eliminating out-of-memory errors and representation collapse. Graph expansion yields 4.7–5.6% token reduction per task, while the incremental sync daemon cuts index update latency by 96.5–99.9% compared to full recomputation. **Limitations:** The approach depends on the quality and completeness of the constructed functional graph; building and maintaining this graph for highly dynamic tool registries may introduce nontrivial overhead. The learned reranker and R‑GCN training require additional computational resources, and evaluation is limited to two benchmark distributions, leaving real-world deployment scenarios unaddressed. Furthermore, graph sparsification risks discarding potentially useful functional connections in edge cases, and the incremental synchronization daemon may face consistency challenges under extreme churn rates.

AI-generated (deepseek-v4-flash) · created 2026-06-09