- Updated: August 26, 2026
- 2 min read
SchemaRouter: Field-Aware Tool Routing for Efficient Heterogeneous Agentic RAG
SchemaRouter: Field-Aware Tool Routing for Efficient Heterogeneous Agentic RAG
Authors: Yong-eun Cho
Published on: August 25, 2026
In the rapidly evolving landscape of retrieval‑augmented generation (RAG), heterogeneous agentic systems must orchestrate a wide variety of external APIs, internal databases, vector stores, and graph stores. Traditional approaches either expose every tool description to the LLM—causing over‑fetching and high latency—or rely solely on vector similarity, which often leads to under‑fetching and incomplete answers.
We introduce SchemaRouter, a lightweight routing layer that models tools, endpoints, parameters, response fields, domain concepts, units, provenance, and license policies as a unified schema graph. Given a user query, SchemaRouter emits an executable tool plan that precisely selects the necessary tools and fields, dramatically reducing token usage and latency while preserving answer accuracy.
Key Features
- Field‑aware routing based on intent‑group projection and concept‑field matching.
- Deterministic field selection via an alias layer for robust tool execution.
- Built‑in provenance and license grounding for transparent, compliant answers.
Performance Highlights
On a materials‑science benchmark (110 queries), SchemaRouter achieved:
- Answer accuracy: 0.71 (matching the “fetch‑everything” baseline within confidence intervals).
- Tool‑exact rate: 0.93.
- Parameter validity: 1.0.
- Token usage: 227 retrieved‑context tokens vs. 2,066 for fetch‑everything.
- Latency: 2.7× faster than the prompt‑all baseline.
- Provenance & license grounding in 62 % of answers.
Why SchemaRouter Matters
Minimizing selected‑field count alone proved counter‑productive, dropping accuracy to 0.56 with negligible token savings. SchemaRouter’s projection strategy preserves recall while delivering efficiency gains, making it ideal for large‑scale, schema‑size‑independent deployments.
Read More
Explore the full paper on arXiv and discover how SchemaRouter can power your next AI‑driven knowledge system. For implementation details, visit our internal resources at ubos.tech/docs/schemarouter.
Stay updated with the latest advancements in AI‑driven retrieval and generation by following our blog at ubos.tech/blog.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.