FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents

FinCacheServe enables dependency-consistent answer reuse for RAG systems by indexing responses using document versions, evidence fingerprints, and model configurations. A vLLM implementation using Qwen2.5 models achieved a 53.31% LLM call reduction on 32B parameter workloads, outperforming versioned semantic caching's 38.97% hit rate while maintaining zero stale outputs.