FinCacheServe Boosts RAG Efficiency with Dependency-Consistent Answer Reuse.

Serving Retrieval-Augmented Generation (RAG) applications over dynamic enterprise documents, such as financial filings, often involves repeatedly executing semantically identical analysis requests. This leads to redundant, GPU-intensive generation work. FinCacheServe addresses this by implementing a sophisticated answer reuse mechanism that maintains dependency consistency even when underlying documents, evidence, or tool outputs change. FinCacheServe treats each generated answer as a serving object, indexed by the enterprise intent and protected by a comprehensive set of identifiers including document versions, evidence fingerprints, tool fingerprints, model identity, and decoding configurations. A vLLM implementation, tested with SEC-derived financial document workloads and Qwen2.5 models, demonstrated remarkable efficiency gains. On a 2,230-request trace, FinCacheServe skipped 53.27% of LLM calls without any observed dependency-stale outputs. Across other operator-suite seeds, it consistently outperformed versioned semantic caching and grounded-style reuse, achieving 44.30% lower estimated energy consumption per fresh, successful response within a 2-second service level objective.