Route a support message
The decision: which team should inspect this message? Provide the message and the categories you operate. Use a Choice with clear boundaries and an “other” option. The Choice documentation explains the underlying question type.
Message: “The export button fails on my monthly report.”
Options: product, payments, other.
Application action: add a suggested team label for review.
For an initial pilot, keep the existing queue unchanged and compare suggestions against the team’s real assignments. Include mixed cases such as “I paid for an export feature that fails.” A label should help the team; it should not conceal the second issue.
Screen retrieved passages
The decision: how well does a passage address a reader’s question? A Score evaluates ordered, descriptive levels. Consult the Score reference for its response fields.
Question: “Can I export data without an account?”
Levels: unrelated, mentions export without answering access, directly answers the access requirement.
Application action: retain useful passages for the answering step.
Our suggested evaluation includes passages that discuss the right product but the wrong feature. Preserve the original source URL and text. A relevance score alone does not certify that the source is current or true.
Flag a request for human review
The decision: does a message ask to delete an account? A Noul evaluates a yes/no proposition and returns a probability. See the Noul reference.
Message: “What happens to my saved projects if I close my account?”
Question: does the customer explicitly request account deletion now?
Application action: route to clarification or a review queue.
This is intentionally different from executing deletion. Test informational questions, explicit requests, quoted text, and changed minds. Keep permissions and confirmation in ordinary application logic. A model score does not provide user authorization.
Turn a sketch into a pilot
These are original design examples, not live demos or measured results. Start with recorded, non-sensitive examples and a clear expected answer. Add edge cases before using the result to change a real workflow.
For each pilot, write down four things: the decision, the evidence supplied, what happens on uncertainty, and how a wrong answer is detected. If those are unclear, refine the workflow before adding another model call.
Use the API walkthrough to connect one question, then review the known limitations.
Sources & verification
Checked September 23, 2026. This guide summarizes documentation; it is not an independent benchmark. Provider details can change.