Start with one refund request
Import the Refund Approval scenario and separate the evidence and roles needed for an AI recommendation and a human decision.
Question for this chapter
Which evidence and roles are needed to review one refund request?
Why this matters now
Starting with the Agent makes a fluent sentence look authoritative. First, separate transaction evidence, risk signal, policy, recommendation, and final decision into distinct assets and responsibilities.
Try it
refund_approval.zip Download(43 KB)In the left sidebar, select Collections. On the Collections page, choose the upper-right more (⋯) menu → Import (가져오기) and upload the ZIP. After validation, a secrets step asks for the Claude API key the Agent will use — you can create the secret right there if you do not have one yet. Bind it, then select Start import.
Success looks like this
The collection tree shows raw_data (Raw Data) and refund_ops (Refund Operations Data). The current
scenario places its data and executable assets under refund_ops; raw_data is empty.
| Evidence or role | Question it answers | Asset to inspect |
|---|---|---|
| Transaction evidence | Who bought what, for how much, with what refund history? | order, customer, and refund datasets |
| Risk signal | Which request should be reviewed first? | risk_assessed_orders |
| Policy evidence | Which return-window, amount, and reason rules apply? | refund_policy knowledge |
| AI recommendation | Which option best fits the combined evidence? | refund-approval Agent |
| Final decision | Who confirms approved, partial, or rejected? | capture_review_decision Actor |
Deep dive — assets imported by the scenario
| Asset type | Count | Contents |
|---|---|---|
| Collections | 2 | raw_data, refund_ops |
| LLM model | 1 | external Claude model; references the secret you bind during import |
| Datasets | 5 | raw orders, customers, refunds, enriched orders, risk-assessed orders |
| Codes / pipelines | 3 each | enrichment, risk assessment, ontology materialization |
| Ontology | 3 entities / 2 relations | customers, orders, refunds, and two relations |
| Knowledge / dashboard | 1 each | refund_policy, refund_operations |
| Agent tools / Actors | 2 tools / 3 Actors | order lookup, policy search, confirmation, downstream actions |
| Agent | 1 | refund-approval |
Next decision
You have separated the roles. Next, attach customer context and past refund history to all 12 orders so a reviewer can read the evidence together.