What the documents look like
Order fulfilment and after-sales answering for a Tmall flagship store draws on four recurring materials: the product publishing page, the brand authorization letter, the delivery commitment record, and the after-sales ticket. Brand authorization and flagship identity make qualification checks a routine action, while the delivery commitment record fixes a time requirement onto each individual order. These materials update daily and per order, so a single conversation may move between product-page information, a committed time limit, and an after-sales ticket. Multi-turn prompts must treat these materials as a defined reference scope and leave a gap when the materials do not cover the question, rather than filling in business facts. The materials differ in subject, object, and validity range, so the prompt has to separate the material itself from any inference drawn about it.
Constraints at the multi-turn prompts step
The reference scope has to be stated. The brand authorization letter, the delivery commitment record, and the after-sales ticket each answer a different question. The prompt should name which classes of material this answer may cite and how to proceed when a class is missing. For history, a new session and a continued session carry different context, so the history content that actually enters the request needs to be checked. For question optimization, continuous knowledge-base conversation allows checking of referent resolution and question expansion; short questions do not all qualify. For model and length, the model's supported context and the reply settings need checking, together with long-document and follow-up combinations. For reply format, choose the format according to what comes next; when json_schema is selected, fill in the corresponding format requirement and verify the returned structure. Retrieval and similarity do not carry permission or entity isolation, and a retrieval result is not an authorization conclusion.
Configuration and implementation checks
The table below lists, for the sample on this page, official documented settings or implementation check items relevant to this feature and primary task.
| Configuration item | Assumed starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| AI model | Assume a model that supports multimodal recognition is selected | The sample contains an image link, so compare usable options against a text-only model |
| Memory rounds | Assume a small number of history rounds as the starting point | Customer service and knowledge-base answering typically keep a small number of history rounds; compare consumption and response against carrying all history |
| Reply cap | Assume it is opened and kept at a medium length | After-sales answering needs a complete explanation; compare against a too-short truncation and a too-long consumption case |
| Temperature | Assume a low value | A rule-defined scenario gives more stable answers; compare against the divergence at a higher value |
| Reply format | Assume the default is kept | Ordinary customer-service answering needs no fixed structure; compare against selecting json_schema and filling in the format requirement |
| Multimodal recognition | Assume image recognition is opened | The sample contains an image link, so confirm the model supports that type; compare against the popup notice when it does not |
Three common errors
- Writing the reference scope as all materials. In the hypothetical exercise sample, the material text available for checking reads: "The brand authorization letter of this store states that the authorizing subject is Company A, the authorized object is the Tmall flagship store, and the authorization is valid until 31 December 2026." "The delivery commitment record shows that order 20260501-001 was committed to ship within 48 hours, and the actual shipment time was 30 hours after the order." "After-sales ticket 20260501-002 has the status pending review, and the application content is a refund." The draft sentence reads: "The order was shipped within the committed time limit, and the refund application has been approved." The error is writing a pending review as approved and merging different materials into one conclusion.
- Treating a retrieval result as an authorization conclusion. A retrieved material only shows that the content is related; it does not show that a subject holds permission, that an order was fulfilled, or that a refund was approved.
- Treating an unstated field as a negative. When a material does not state a field, record it as unknown or pending verification; do not write it as absent or conflicting.
How to verify the setup
- Input the three material sentences above together with the draft sentence, and compare whether the draft sentence writes "pending review" as "approved" and whether it merges the delivery time limit with the refund status into one fact. If the answer keeps the pending review status and states that the refund has not been approved, the condition is met; if it writes approved directly, the condition is not met.
- Input a continued-session request under the same chatId, and compare the history content actually carried in against the configured memory rounds. If the carried rounds match the setting, the condition is met; if all history is carried in or none is carried in, record the difference.
- Input a question containing an image link, and compare whether the model supports image recognition and whether that option is opened. If the model does not support it, the popup should show the unsupported notice; if it supports it and the option is open, the image content enters the request. If the returned content is unrelated to the image, record it as unsupported rather than inferring a permission or entity filter result.
The hypothetical exercise sample for this page is: material text covering a brand authorization letter with a stated subject, object, and validity end date; a delivery commitment record for one order with a committed time limit and an actual shipment time; and an after-sales ticket with a pending review status and a refund application. The draft sentence to be judged is the one given in the first error above. A pending review is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Missing evidence stays missing, and retrieval is not authorization.