Steer

Turn Observe’s quality scores and your feedback into better few-shot examples and prompts.

Steer closes the loop on quality. It takes the scores Observe writes on your calls, plus the thumbs-up/down feedback your app sends, and uses them to pick the strongest few-shot examples and optimize your prompts. The more calls Observe scores, the better the signal Steer works from.

How it works

• Quality scores. Every call an Observe rule judges gets a score and observation stored against it.
• Feedback. Feedback you send from your SDK attaches a human signal to a call, alongside the judge’s scores.
• Better outputs. Steer uses that signal to select few-shot examples from your top-scoring calls and to tune the prompt, so quality climbs as more calls come in.

Requires Observe

Steer runs on the scores the Observe judge produces, so you need at least one Observe rule scoring the calls you want to improve. Once it’s scoring, open Controls → Steer to manage it.

Score the calls you care about with Observe and send feedback from your SDK. The more signal Steer has, the better the examples and prompts it picks.