What the documents look like
Refund-only disputes differ from ordinary return-and-refund cases in one structural way: there is usually no returned parcel entering the warehouse. The evidence that can be checked therefore concentrates on dispatch, delivery acceptance, and the value of the goods themselves. In store service and internal training, the material for each case comes from the platform after-sales backend, carrier tracking, customer service chat logs, and warehouse outbound records. Each order generates a fresh set of documents, and the sets are summarized daily. The multi-turn prompts step has to organize that material into a traceable context, keep the model answering inside the scope of the supplied documents, and leave gaps visible when the evidence is incomplete. It does not decide whether a return is required, and it does not submit an appeal.
Constraints at the multi-turn prompts step
In the chat interface, an empty chatId means the context function provided by FastGPT is not used at all, and the context is built entirely from the messages that are passed in. A non-empty chatId means the session is retrieved from the database automatically, the last item in the messages array is taken as the user question, and the remaining messages are ignored. Continuing a session and starting a new session therefore carry different history, and any check has to distinguish the two. Memory turns control how many earlier rounds of chat the AI may reference at most; customer service applications usually keep a small number of history rounds. Reply limit, temperature, Top_p, stop sequences, and reply format are existing node behaviors or fields that can actually be configured. Retrieval on its own does not constitute authorization. A form collects values; business verification, approval, database writes, and external synchronization each need their own downstream implementation. Similarity and relevance scores do not carry permission or entity-isolation duties.
Hypothetical drill sample: a store after-sales specialist prepares a handover request for a supervisor during internal training. The source material reads: "Order A was signed for by the carrier on March 12; the signature field is blank. The warehouse outbound record shows one unit of the same product left the warehouse on March 10. The buyer applied for a refund only, on the ground that the goods were not received." The proposed sentence reads: "This order already has proof of delivery; handle it as delivered and reject the refund-only application." What must be judged is whether the proposed sentence is supported by the material, and what authorization state the handover request is in.
Configuration and implementation checks
The table below is built around that sample and lists officially described settings and implementation checks relevant to this feature and its primary task. The values are starting points for this hypothetical trial, used to compare against another set of settings, and are not official defaults.
| Configuration item | Hypothetical trial starting point and reason | Check method |
|---|---|---|
| AI model | Choose a model that supports the needed context length and multimodal capability; the specific capability shown in the configuration dialog governs | Open the AI configuration dialog and read the displayed maximum context, tool calling, and multimodal capability |
| Memory turns | Start with a small number of history rounds, because customer service question answering usually keeps a small history | Compare against a setting that keeps no history, and confirm the carried range |
| Reply limit | Start with a length that can hold a complete explanation, because a check conclusion must be output | Compare against a setting low enough to truncate |
| Temperature | Start low, because a rule-defined scenario needs stable answers | Compare against a higher setting for answer stability |
| Reply format | Keep the default, because no later step reads a fixed structure | If changed to json_schema, fill in the corresponding format requirement and check the returned structure |
| Hide AI output | Decide by whether a later node must rewrite the content | After enabling, check what the user sees against the input to the later node |
Three common errors
- Treating an empty chatId and a non-empty chatId as the same context range. The input is the same messages array under both chatId states; compare the history actually carried, and observe whether history is ignored or the last item is taken as the only question.
- Reading a blank signature field as proof of delivery. The input is the source material above; compare "the signature field is blank" with "already has proof of delivery" in the proposed sentence, and observe whether the proposed sentence is supported by the material.
- Treating a handover request awaiting submission as authorized execution. The input is the proposed sentence and the request state; compare the stage the request is in against actions such as rejection or submission, and observe whether a pending review is treated as approved.
How to verify the setup
- For the sample source material, input that passage together with the proposed sentence, and compare "already has proof of delivery" in the draft against "the signature field is blank" in the source. The expected observation is that the draft is not supported by the material and must remain pending verification. If the model directly concludes that delivery was accepted, that round does not meet the requirement.
- For continuing a session, input the same chatId with one new question, and compare the history rounds carried against the memory turns setting. The expected observation is that the carried range matches the setting. If the carried range exceeds the setting, record the actual observation without inferring a cause.
- For empty-result handling, input paper messages together with a retrieval result of [] that clearly returns no content, and compare how the downstream node is designed to output under an empty reference. The expected observation is that the downstream handles it as an empty reference; do not claim that a fictional rule retrieval necessarily returns []. Libraries built on different vector models cannot be associated together; check the identifier first, exclude those that do not match, and record the actual observation.