What the documents look like
Promotion enrollment and price management in this category revolves around four working documents: the campaign enrollment form, the promotional price approval sheet, the inventory lock sheet, and the campaign review sheet. In store service and internal training, multi-turn dialogue carries a specific job: it lets store staff check material scope, price wording, and approval status across a series of follow-up questions. The prompt must state the document scope, the answering task, and how to handle missing evidence. Compared with product-title compliance, the core data here is price, inventory, and campaign rhythm; advertising wording review is not the main object.
The main data sources are the Tmall marketing backend, the JD campaign backend, ERP inventory, and historical transaction prices, updated weekly or per campaign. A frequent training problem is that the material states one campaign's enrollment price while the follow-up question concerns a different campaign or channel. Multi-turn dialogue then has to distinguish a new session from a continued session and check which context was actually carried in. Content outside the material scope is marked as not provided, and the model does not supply business facts on its own.
Constraints at the multi-turn prompts step
The first constraint comes from history records. When chatId is passed, the session is taken from the database and the last item in the messages array is used as the user question, while the remaining messages are ignored. When chatId is not passed, context is built entirely from the supplied messages. Store training has to verify the context actually carried in and separate a new session from a continued session.
Where continuous knowledge-base dialogue is needed, check coreference resolution and question expansion rather than assuming every short question qualifies. For model and length, check the context and reply settings the model supports, and test long documents combined with follow-ups. Reply format follows the downstream consumer; when json_schema is selected, fill in the corresponding format requirement and validate the returned structure. Retrieval and similarity do not carry permission or entity isolation, and the same-embedding-model requirement for a knowledge base does not extend to other external systems.
Hypothetical paper exercise
Material text for checking: (1) The "East China Region June Store Promotion Enrollment Form" states the enrolling entity is the Hangzhou West Lake store, the campaign object is in-store item A, the enrollment price is 199 yuan, the enrollment period is June 1 to June 30, and the form status is "pending regional review." (2) The "Promotional Price Approval Sheet" states item A's promotional price is 199 yuan, the approval status is "submitted, pending finance BP review," and no effective date is stated.
Draft sentence to be judged: "Item A's 199 yuan promotional price has been approved and can be executed from June 1."
This sentence turns "pending finance BP review" into "approved" and turns an unstated effective date into a definite date. The approval sheet being pending review does not constitute completed approval, and the enrollment form being pending regional review does not constitute enrollment acceptance. Both evidence gaps remain open.
Configuration and implementation checks
The table below lists, for this exercise, official documented settings or implementation checks tied to this feature and task. The values are hypothetical starting points for this paper exercise, used for comparison against the sample material, and are not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Object and comparison method |
|---|---|---|
| AI model | Select a model whose dialog shows support for the needed context length | Compare with the material length and the follow-up combination, confirming the model's maximum context can hold them |
| Memory turns | Start at 2 turns | Compare with the store's follow-up turns, confirming the previous turn's material wording can be carried in |
| Reply limit | Start at a length that can fully output the checking conclusion | Compare with the draft sentence length, confirming the conclusion is not truncated |
| Temperature | Start at a low value | Compare with the rule-clear checking scenario, confirming stable answers |
| Reply format | Keep default at the start | If a downstream flow needs a fixed structure, switch to json_schema and fill in the format requirement |
| Multimodal recognition | Off at the start | Compare with the material type; plain text material needs no activation |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the enrollment form status is "pending regional review"; retrieving that form does not mean enrollment was accepted, nor that the price was approved. The input is the sample material text, the comparison object is the status field stated in the material, and the observable result is whether the answer writes pending review as accepted.
- Writing pending review as completed approval. The approval sheet states "submitted, pending finance BP review," while the draft sentence writes "has been approved." The input is the approval sheet text and the draft sentence, the comparison object is the approval status field, and the observable result is whether the two agree.
- Filling an unstated effective date with a definite date. The approval sheet states no effective date, while the draft sentence writes "can be executed from June 1." The input is the approval sheet text, the comparison object is whether an effective date is stated, and the observable result is whether the answer invents a date.
How to verify the setup
- Input the sample material text and the draft sentence, compare the approval status in the draft sentence against the "submitted, pending finance BP review" stated in the approval sheet, and observe whether the answer points out the mismatch. The expectation is that it points out the mismatch and keeps the item pending verification; no actual test is claimed.
- Input the sample material text and the follow-up "was this campaign's enrollment accepted," compare against the "pending regional review" stated in the enrollment form, and observe whether the answer explains that the enrollment status is pending and does not constitute acceptance. The expectation is that it states the item is pending verification.
- Input the sample material text and the follow-up "what is the approval sheet's effective date," compare whether the approval sheet states an effective date, and observe whether the answer explains that the material does not state this field. The expectation is that it records the evidence gap.
All three checks use the hypothetical text and draft sentence above, each stating the input, the comparison object, and the observable phenomenon. If retrieval returns an empty result, the downstream design can be checked against the explicit paper input [], without claiming that a fictional rule search necessarily returns empty.