ArisGlobal cut Amazon Bedrock inference spend ~40% by caching the shared prompt prefix
ArisGlobal's NavaX pharmacovigilance engine extracts ~130 fields from adverse-event PDFs using sequential Claude requests on Amazon Bedrock, with a shared prompt prefix that grew from about 72,000 to 180,000 tokens over six months as tenants were added, pushing per-case inference costs up more than 200%. The original design cached the per-case source document first, which left the growing shared prefix uncached and reprocessed at full input rate on every request. Reordering the request so the shared prefix precedes the cache checkpoint made the cache entry reusable across requests, cases and tenants. ArisGlobal reports 65% of input tokens served from cache, roughly 50% lower effective input cost, about 40% lower monthly inference spend at over 19 billion tokens per month, up to 25% lower latency, and around 15% more tenants onboarded without proportional cost growth.