answer-caching

FinCacheServe is a system for dependency-consistent answer reuse in RAG serving over mutable enterprise documents, using document versions, evidence fingerprints, and tool fingerprints to invalidate caches. Evaluations show it skips over 53% of LLM calls with zero stale outputs, reducing GPU cost compared to versioned semantic caching.