What the documents look like
Refund and return handling on a general marketplace runs per order and continues throughout the order lifecycle. It does not depend on a single live-stream commitment, so the same order can generate a refund request, a return logistics record, quality-inspection photos and a refund approval at different times. Four sources carry the operational state: the platform after-sales backend, the customer-service chat, the WMS receiving record, and the payment refund record. Each source holds its own fields and updates at its own pace. The brief states the update frequency as per order and daily; this page does not derive batch synchronization, full synchronization, real-time or transaction-triggered mechanisms from that statement.
A multi-turn prompt that treats one system's current value as the conclusion for the whole chain will drift on responsibility judgment and refund approval. The prompt needs to state the material scope, the answering task, and the handling of missing evidence, so the model does not fill in business facts. When a conversation continues, the interface reads the session from the database using chatId, takes the last item of the messages array as the user question, and ignores the remaining messages. When chatId is not passed, no context is taken from the FastGPT database and the full messages array builds the context. Retrieval and similarity decide which fragments are returned; they do not carry permission or entity isolation. Reply format is chosen by the downstream consumer, and selecting json_schema requires the corresponding format specification to be filled in.
Hypothetical paper exercise. A customer-service agent works order A-7731 in a session with chatId H-2024-0912. The material text reads: "Refund application A-7731 states application date 2024-09-10, requested amount 128, status pending review." "Return logistics record shows delivery signed on 2024-09-11, signature note states outer packaging damaged." "Three quality-inspection photos, capture date 2024-09-11, no inspection conclusion attached." The proposed draft sentence is: "The return for this order has been signed for and inspection shows no abnormality, so the refund can be executed." The material does not state an inspection conclusion and does not state an approval result. The draft sentence's claims that inspection shows no abnormality and that the refund can be executed are both unsupported.
Constraints at the multi-turn prompts step
The prompt must define the material scope, the answering task and the missing-evidence rule before any answer is drafted. In this category the material carries status values such as pending review, and it carries logistics events such as signed. A signed return is not an inspection pass, and a pending-review application is not an approval. The prompt therefore needs a rule that separates a recorded event from a completed business decision.
Session handling adds a second constraint. Continuing a session with a chatId means the last message is the user question and earlier messages in the request are ignored. Starting without a chatId means the request's messages build the entire context. A prompt that assumes prior turns are always present will behave differently across the two paths, so the setup must state which path is being used.
Retrieval returns fragments ranked by similarity. A high-similarity refund fragment is still only a returned fragment; it does not authorize execution, and it does not filter by category, entity or permission. The prompt must not treat a returned fragment as an approved action. Reply format is a downstream contract: when the next step reads a fixed structure, json_schema is selected and its format requirement is supplied; otherwise the default is kept.
Configuration and implementation checks
The table below ties each row to the hypothetical exercise above. Values are hypothetical starting points for this exercise, not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model whose configuration dialog shows a maximum context long enough for the sample material plus follow-up | Compare the model's maximum context against the total length of the sample material and the follow-up question; price alone is not the basis |
| Memory turns | A small number of history turns | Customer-service applications typically retain a small number of history turns; compare the historyPreview actually carried into the continued session against the configured number |
| Temperature | A low value | Rule-bound after-sales answers suit a low value; compare whether the same material produces a stable answer across values |
| Reply format | Default | The downstream step does not read a fixed structure; compare whether the returned content can be consumed directly |
| Multimodal recognition | Off | The sample quality-inspection photos carry no conclusion; compare whether enabling it makes the model read the image and whether the conclusion field is still absent |
| Hide AI output | Off | Debugging needs the intermediate result visible; compare the user-visible content and the downstream node input before and after the switch |
Point values, maximum context, tool calling and multimodal capability follow what the configuration dialog actually displays. A model that does not support multimodal recognition does not show that option; a model that does not support thinking configuration does not show those options either.
Three common errors
- Treating a retrieval hit as authorization to execute. Input: the sample material and the proposed draft sentence. Compare: whether the draft sentence contains an approval conclusion that the material does not state. Observable result: the answer preserves the pending-review status and does not assert execution. The error appears when the answer endorses execution from a high-similarity fragment.
- Reading pending review as approved. Input: the refund application text. Compare: the status field value against the conclusion in the draft sentence. Observable result: the answer keeps the status as pending review. The error appears when the answer states the application is approved.
- Reading a signed return as a passed inspection. Input: the return logistics record and the quality-inspection photo description. Compare: whether an inspection conclusion field exists in the material. Observable result: the answer records the conclusion as absent. The error appears when the answer supplies a conclusion the material does not carry.
How to verify the setup
- Input the sample material and the proposed draft sentence. Compare the phrases "inspection shows no abnormality" and "the refund can be executed" against the material fields. The check is satisfied when the answer states both claims are unsupported and preserves the pending-review status. The check is not satisfied when the answer endorses the refund directly.
- Input the paper input [] for the empty-result handoff. Compare the downstream node's design for the case with no returned fragment. The check is satisfied when the downstream step does not treat the empty result as a verified conclusion. The check is not satisfied when the downstream step produces a business conclusion from the empty result. This is a paper check; it does not claim that a fictional rule retrieval returns [].
- Input a continuation request in the same session. Compare the history turns actually carried in against the configured memory turns. The check is satisfied when the carried range matches the configuration. The check is not satisfied when the continuation request omits the required prior context or carries a range that differs from the configuration. The retrieval check observes returned fragments and the failure branch only; it does not guarantee a first hit or full recall. Libraries across different vector models cannot be associated together, so identifiers are checked first and non-matching libraries are excluded.
These checks describe expected behavior and do not claim that any test, deployment, approval, refund or other business action has occurred. Where the material does not state a field, the gap remains a gap.