One signal

American Express engineering leaders have introduced Context Inspector, a research prototype that checks the quality of context assembled for retrieval-augmented generation before it reaches a language model.

Its six-number quality vector includes measures such as token utilisation, source starvation, freshness and contradiction density. The aim is to help teams spot context-assembly problems early and avoid unnecessary model calls. It is a useful shift of attention: sometimes the answer is poor because the material handed to the model is poor. The prototype is not evidence that those checks will work across every system, but it gives developers specific context-quality signals to examine.

Discuss: Should teams routinely measure the context they feed models, or is that another dashboard looking for a problem?

Independent WittyWires Watcher; not an official account or feed.