What the documents look like
Copy drafting for a US TikTok Shop listing draws on four record families: Seller Center listing fields, order-center return and refund records, Return Grading evidence images, and settlement bills. The brief marks these as updated daily and per order, so a single drafting conversation can reference material captured at different times for the same product. Return windows and return-freight wording that appear in a draft must match what the underlying records state; the brief lists platform automatic refunds, higher return fees, and failed appeals as consequences of getting this wrong.
The multi-turn task is narrower than the record set. It is to keep a listing draft, an evidence description, and a settlement wording aligned sentence by sentence across consecutive follow-up turns. The records supply the material identities; the conversation supplies the alignment work. Nothing in the brief authorizes the model to add business facts that no record contains.
Constraints at the multi-turn prompts step
History handling is documented. When a non-empty chatId is passed, FastGPT takes the session from its database and uses the last item of the messages array as the user question; the remaining messages are ignored. When chatId is not passed, no context is used and the context is built entirely from messages. A prompt therefore has to state the material scope, the drafting task, and how to proceed when evidence is missing. The model must not fill in business facts.
Reply format follows the downstream consumer. Selecting json_schema requires filling in the corresponding format requirement and verifying the returned structure. Model and length settings need checking against the context the model supports and the reply length the draft requires. For knowledge-base multi-turn use, question optimization covers reference resolution and question expansion.
Hypothetical drill sample, labelled as such and not drawn from any real record. Source text for comparison: "The return window for this product is 30 days; when the return reason is wrong size, the buyer bears the return shipping cost. The Return Grading record shows one order on 2024-10-05 graded as minor wear." Draft sentence to be judged: "This product supports 30-day no-reason returns with return shipping fully free." The comparison object is the source text against the draft sentence. The source states a shipping-cost allocation only for the wrong-size case; it does not state that shipping is free in all cases, and it does not state "no reason." The draft sentence goes beyond the source scope and is judged unsupported.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Relation to the sample |
|---|---|---|
| AI model | Select a model supporting the required context length and multimodal ability | Image recognition is needed when the sample includes an evidence-image description |
| Memory turns | Set the number of prior turns to reference according to the follow-up need | A small number of turns is retained when checking listing and evidence across turns |
| Reply cap | Set according to the draft length required | The sample draft sentence is short; too low a cap truncates it |
| Temperature | A higher value suits copy drafting | The sample judgment needs stability, so a lower value is appropriate |
| Reply format | Choose json_schema when a fixed structure is needed and fill in the format requirement | The sample judgment output can use a fixed structure |
| Multimodal recognition | Enable the corresponding type when the model supports it | The sample Return Grading evidence image needs image recognition |
The official documents do not give default values for these numbers, and this page does not specify any. Where a number is needed, treat it as a starting value for this hypothetical trial, state why it was chosen, and state what value it is compared against.
Three common errors
- Writing "return shipping fully free" when the source says the buyer bears return shipping for the wrong-size case. Input is the source text; comparison is the shipping-cost allocation field; the observable result is that the draft sentence and the source text take different values in the same field.
- Writing "the product has no quality problem" when the Return Grading record says "minor wear." Input is the evidence description; comparison is the wear-state field; the observable result is that the draft sentence changes the state value recorded in the source.
- Writing "the platform has confirmed the refund" when the settlement bill records a deduction. Input is the bill record; comparison is the deduction and refund status fields; the observable result is that the draft sentence replaces a recorded deduction with a refund fact.
How to verify the setup
- Input the sample source text and the draft sentence, compare the shipping-cost allocation field, and judge the draft unsupported when it says "shipping fully free" while the source states the allocation only for the wrong-size case.
- Input the Return Grading evidence description and the draft sentence, compare the wear-state field, and judge the draft unsupported when it says "no quality problem" while the source says "minor wear."
- Input the settlement bill record and the draft sentence, compare the deduction and refund fields, and judge the draft unsupported when it says "refund confirmed" while the source records only a deduction. If retrieval returns an empty result, a clearly labelled paper input of
[]can be used to check the downstream design; this page does not claim that a search over the hypothetical rule would return[].
Retrieval and similarity scores do not perform permission or entity isolation, and a shared vector model requirement does not extend to external systems. Where the documents do not state an error code or an observed outcome, record what to compare and leave the rest unknown.