GroundedCache: Evidence-Validated Routing for Safe Answer Reuse in RAG
A recent publication on arXiv introduces GroundedCache, a cache router designed for retrieval-augmented generation (RAG) systems that is validated by evidence. The authors highlight that the primary issue with output-level semantic answer caches is safety rather than speed. They note that similar prompts can lead to various correct answers, evidence can change as the corpus evolves, and adversarial attacks may manipulate cached responses. GroundedCache allows a cached answer only when four inexpensive gates are satisfied: query similarity, overlap of retrieved evidence, validity of the source version, and lexical (or judge-based) consistency. This system seeks to minimize token costs and time-to-first-token (TTFT) while maintaining accuracy. The paper can be found on arXiv under ID 2605.27494.
Key facts
- GroundedCache is an evidence-validated cache router for RAG systems.
- It addresses safety issues in output-level semantic answer caches.
- Four gates must hold: query similarity, retrieved-evidence overlap, source-version validity, and lexical consistency.
- The paper argues that safe reuse is more important than fast reuse.
- Adversarial collision attacks can hijack cached responses.
- Retrieved evidence drifts as the corpus is updated.
- The system aims to reduce token cost and TTFT.
- Paper available on arXiv with ID 2605.27494.
Entities
Institutions
- arXiv