What the documents look like
Retail checkout and loyalty work runs on four document families: promotional price tags, membership points rules, the cashier daily report, and exception refund slips. The stated sources are the POS system, membership CRM, payment platform, and price tag system. The brief records the update rhythm as daily or per shift. Store cashier supervisors, CRM operations staff, and finance auditors share the same conversation surface, so a multi-turn prompt has to keep promotional prices, points entitlements, and refund wording inside the material actually supplied. The stated consequences of error are price complaints, cash shortfalls or overages, and membership entitlement disputes.
For this page the relevant difference from convenience-store fresh-food ordering is that checkout promotion work concerns price and membership benefits, and does not handle short-shelf-life inventory. That shapes the prompt: the conversation must state the document scope, the answering task, and how to behave when evidence is missing, so the model does not write in business facts that the source material never carried.
Constraints at the multi-turn prompts step
Multi-turn behavior depends on which context is actually carried in. When a chatId is supplied, the system takes the session from the database and uses the last element of the messages array as the user question, ignoring the remaining messages. When chatId is not supplied, context is built entirely from the submitted messages. These two modes produce different inputs for the same follow-up question, so a training prompt has to state which mode is in use.
Model and length settings also bound the conversation. The maximum context figure describes how much content a single request can reference; a larger value suits long documents and long dialogues. The reply cap, when enabled, limits reply length, and a low setting can truncate an answer. Reply format matters when a downstream step reads fixed structure; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure. Memory rounds control how many earlier rounds the model may reference. The documented question-guide interface, from version 4.8.16, requires appId and chatId and pulls the most recent six rounds of dialogue as context. That is a documented condition of that interface, not a general default for every conversation.
A hypothetical paper exercise sample
Material text supplied for comparison (paper exercise, not a real record):
"From 1 March 2025 to 7 March 2025, the membership points rule at Store A is one point per 1 yuan spent, and points may be deducted against the next purchase. The cashier daily report for 3 March shows one exception refund slip at Store A, refund amount 58 yuan, status pending supervisor review."
One draft sentence to be judged:
"The refund slip has passed supervisor review, and the membership points have been synchronised and deducted."
The draft sentence concerns subject Store A, object the 3 March exception refund slip, status pending supervisor review, and date range 1 March to 7 March 2025. The authorisation boundary is this: pending review is not approved review, the material does not record any points deduction action, and the draft sentence exceeds what the material supports.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Hypothetical starting value and reason |
|---|---|---|
| Memory rounds | Controls how many earlier rounds the model may reference; customer-service style use keeps a small number | Paper starting value 6, matching the documented six-round context pull of the question-guide interface, so follow-up carry-over can be compared |
| Reply cap | When enabled, limits reply length; a low value can truncate | Paper starting value set to hold the material text plus the draft sentence, compared against truncation |
| Temperature | Lower values give steadier answers and suit rule-bound scenarios | Paper starting value set low, compared against a more divergent answer |
| Reply format | Ordinary question answering keeps the default; fixed downstream reading requires json_schema with the format requirement filled in | Paper starting value kept at default, compared against a fixed-structure requirement |
| Multimodal recognition | Enabled by type when the model supports it | Paper starting value off, since this sample is plain text, compared against an image-input scenario |
| Thinking configuration | Shown for some models; controls thinking intensity | Paper starting value kept at default, compared against the documented handling of reverting to default after an error |
Retrieval is not authorisation. A knowledge-base search returning similar content does not show that the caller may execute a refund or adjust points. The similarity score is not a category, entity, or permission filter. The thinking budget is a thinking budget and must not be bound to a reference-length limit.
Three common errors
- Writing a pending status as approved. Input: the material text and the draft sentence above. Comparison: the status field. Observable result: the draft sentence asserts a completion action the material does not record, such as "has passed supervisor review" or "points have been synchronised and deducted."
- Treating a retrieval result as authority to act. Input: a retrieval result and an operation request. Comparison: the authorisation boundary. Observable result: an unauthorised business action appears in the answer, such as confirming a refund or a points adjustment.
- Binding the thinking budget to the reference-length limit. Input: the thinking configuration item and the reference-length setting. Comparison: the purpose of each. Observable result: the reference cap is tied to the thinking budget value.
How to verify the setup
- Using the hypothetical sample, input the material text and the draft sentence, then compare whether the draft sentence introduces wording the material does not carry, such as "has passed supervisor review" or "points have been synchronised and deducted." The condition is met when the answer keeps the pending status and states that the points deduction lacks evidence; it is not met when the answer confirms approval.
- For a continuing session, input a follow-up under the same chatId and compare the number of context rounds actually carried against the memory-round setting. The condition is met when the follow-up can reference earlier content without exceeding the set rounds; it is not met when more rounds are carried or none are.
- For empty-result handling, input a paper empty reference list and compare whether the downstream design treats it as no reference. The condition is met when the answer states that no relevant material was found and preserves the evidence gap; it is not met when the answer writes in business facts.