Ground the recommendation in retrieved policy

Question for this chapter

Which policy evidence should be found before the AI recommends a refund decision?

Understand RAG in this situation

RAG prevents the model from relying only on memory. It retrieves relevant passages from refund_policy and places them in the response context. A high-value laptop request may need the 30-day window, the over-$500 escalation rule, and the high-value electronics condition together.

Try it

If model credentials are configured, start a new session with the refund-approval Agent and enter a clearly labeled learning request:

Review a refund for order ORD-DEMO-007.
The reason is defective and the requested amount is $650.

The Agent uses two tools:

1. get_enriched_order — returns order, customer, and risk context
2. search_refund_policy — retrieves policy passages for the reason and amount with RAG

Policy evidence can contain both auto-approval and escalation rules. A request over $500 meets a current human-review condition.

Confirm the fixed learning response

The current get_enriched_order does not query risk_assessed_orders. It preserves only the submitted order_id and returns this fixed learning context for every session:

The ORD-DEMO-007 session is therefore a runtime demonstration, not a lookup that verifies the 12-row batch output from the previous chapter.

Next decision

You have order context and policy evidence for an AI recommendation. Next, confirm that no downstream Actor runs before a person approves the decision.