From the source
- ArisGlobal cut the cost of running its NavaX adverse-event engine on
Amazon Bedrock by caching the shared prompt prefix instead of each source document.
- NavaX extracts around 130 fields from adverse-event PDFs, often 10–200 pages, through multiple sequential model requests that each carry system instructions, per-tenant rules, an extraction schema and the source document.
- As pharmaceutical tenants added compliance requirements, the shared prefix grew from roughly 72,000 to 180,000 tokens in 6 months, a 150% increase, and per-case inference costs rose by more than 200%.
- ArisGlobal had cached the source document first, which left the growing prefix outside the cache; reversing the order reduced effective input cost and let it support more tenants without proportional cost growth.
- ArisGlobal productionized its first generative AI solution on Amazon Bedrock with
Anthropic's Claude in March 2025 and says internal telemetry shows over 1 million real transactions processed in regulated pharmaceutical environments.
- Amazon Bedrock prompt caching is prefix-based, so changing an earlier section invalidates later matches; supported Claude models allow up to four explicit checkpoints, with a default 5-minute TTL and a 1-hour option on select models.






