Evaluate the whole service
A model that performs well in a demonstration may fail on your documents, languages or business exceptions. The NIST Generative AI Profile offers a voluntary framework for considering risks throughout the lifecycle. It is not a certification.
We recommend building a representative evaluation set before launch. Include straightforward requests, conflicting documents, missing information and attempts to push the service outside its scope. Assess whether it can acknowledge that it does not know.
Version what changes the answers
The model is only one part of the system. Record instructions, parameters, available tools and retrieval data versions too. Each change should be traceable to differences in quality, cost or latency. For agents, Anthropic discusses evaluations that account for trajectories and outcomes.
Prepare for operations
Our advice is to set understandable alert thresholds and name the person who can suspend the service. A manual fallback or rollback to a previous version avoids improvisation during an incident. Do not automatically retain every exchange: define what diagnostic information is needed and how long it should be kept.
Sources and features may change. Check the current version when planning your project.
All articles ↗