SchemaRouter: Field-Aware Tool Routing for Efficient Heterogeneous Agentic RAG
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
SchemaRouter cut retrieved-context tokens from 2,066 to 227 while matching fetch-everything accuracy on a 110-query materials-science RAG benchmark and reducing end-to-end latency 2.7x versus prompt-all. The practical takeaway is that routing at the response-field/schema level, not just tool level or vector similarity, can materially lower agentic RAG cost and latency without sacrificing answer quality, while also enforcing provenance/license constraints that baseline tool prompting missed.
SchemaRouter reduces RAG token usage by 9x (227 vs 2,066 tokens per query) while maintaining equivalent accuracy (0.71 vs fetch-everything), enabling production deployments to cut costs and latency without sacrificing answer quality. This matters because it solves the critical tradeoff between precision (under-fetching) and efficiency (over-fetching) in agentic systems, allowing engineers to scale complex tool orchestrations without ballooning inference budgets.