Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer?

GroundedCache proposes a safety-first framework for reusing cached answers in retrieval-augmented generation systems by validating four conditions before serving cached responses. The system achieves near-zero unsafe-served rates (0-1.5%) across benchmarks while maintaining minimal latency overhead, addressing critical vulnerabilities in current caching approaches that can serve incorrect answers.

Caching is fundamental to modern LLM deployments, reducing costs and latency through token reuse. However, existing caching strategies focus primarily on speed optimization rather than safety. GroundedCache reframes the problem: instead of asking how fast answers can be cached and retrieved, the research asks when caching is actually safe. This distinction matters because semantic answer caches in RAG systems face three compounding risks—query variations that map to different correct answers, corpus updates that invalidate cached evidence, and adversarial attacks designed to exploit cache collisions.

The system's innovation lies in its four-gate validation mechanism: query similarity assessment, evidence overlap verification, source-version tracking, and lexical support validation. These checks run with minimal computational overhead while providing defense-in-depth. Testing across HotpotQA and mtRAG datasets using real LLM generations revealed dramatic safety improvements. On HotpotQA, unsafe-served rates dropped from 15-35% to 0%, while mtRAG document-drift scenarios improved from 51.5% to 1.5%. The latency penalty remains negligible at 1.04-1.07x baseline performance.

For LLM infrastructure developers and operators, this research addresses a genuine production concern. As RAG deployments scale and caching becomes standard across serving stacks like vLLM, the risk of silently serving incorrect cached answers compounds. GroundedCache provides an open-source solution with measurable safety guarantees. The ablation studies isolate the lexical support gate as the primary safety mechanism, suggesting simplified implementations may still be effective. Organizations running cost-sensitive RAG applications must evaluate whether their current caching strategies validate answer safety, as the performance gap between unsafe and safe caching narrows considerably.

• →GroundedCache achieves 0% unsafe-served rate on HotpotQA by validating cached answers against four safety gates before reuse.
• →Document corpus drift reduces unsafe-served rates from 51.5% to 1.5%, showing robustness against real-world data changes.
• →Latency overhead remains minimal at 1.04-1.07x compared to uncached baselines, proving safety doesn't require sacrificing performance.
• →Lexical support validation emerges as the load-bearing safety mechanism, with other gates providing secondary defense at near-zero cost.
• →Open-source implementation enables widespread adoption of safety-validated caching across RAG production deployments.