What the documents look like
TikTok Shop US operations material for this category centres on product listings, return and refund records, Return Grading evidence, and settlement statements. Updates arrive daily and per order. Store service and internal training work is multi-turn: a support agent submits a return record screenshot, asks whether a particular deduction belongs to the same order, then requests a short explanation for a new colleague. Constraints come from local return windows and overseas logistics. Materials carry stage states; a return record may sit at pending review, and a settlement statement may not yet exist. Each turn needs to be tied to a specific document rather than carried over from the previous turn.
A hypothetical paper exercise sample: an internal training lead submits a return record whose text reads "Order number US-2024-0871, return request date 6 May 2024, reason: size mismatch, current status is pending review, evidence: three product photos, reviewer has not signed." The same submission includes a draft sentence: "This order's return has been completed and the return fee has been deducted from the settlement statement." The handover request to assess is: "Please use the material above to draft a return-handling note for a new colleague and confirm the fee has been deducted."
Constraints at the multi-turn prompts step
Prompts need to state the material scope, the answering task, and how to behave when evidence is missing, so the model does not fill in business facts. For history, verify what context is actually carried in and distinguish a new session from a continued one. In a continued session, chatId is a non-empty string; the system retrieves the session from the FastGPT database and uses the last item in the messages array as the user question, ignoring the remaining messages. For question optimisation, knowledge-base continuous dialogue needs checking for reference resolution and question expansion, without assuming every short question qualifies. For model and length, check the model's supported context and reply settings, and test long documents combined with follow-up questions. Reply format follows downstream consumption; when json_schema is chosen, fill in the corresponding format requirement and verify the returned structure. Retrieval or forms only collect values; business verification, approval, database writes, and external synchronisation each need their own downstream implementation. A similarity score does not carry permission or entity-isolation duties.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model with multimodal recognition | The material includes product photos; compare the model's capability description with the multimodal options shown in the configuration dialog |
| Memory turns | A small number of history turns as a hypothetical start | Follow-up questions need prior context; compare against the default and adjust to the actual follow-up length |
| Reply cap | A medium length as a hypothetical start | The training note needs complete paragraphs; compare against observed truncation and adjust |
| Temperature | A low value as a hypothetical start | Return-status wording needs stability; compare against a higher setting's more divergent answers |
| Reply format | Keep default unless downstream needs a fixed structure | If a later node reads fixed fields, switch to json_schema and fill in the format requirement |
| Multimodal recognition | Enable image recognition | The material includes product photos; confirm the model supports image input before enabling |
Three common errors
- Writing a pending-review status as completed. Input: the return record text above. Compare "current status is pending review" with the draft sentence "return has been completed." Observe whether the draft treats pending review as approved. Pending review does not constitute completed approval.
- Treating the draft's fee deduction as having occurred. Input: the sample where the settlement statement is absent. Compare the draft sentence "the fee has been deducted from the settlement statement" with whether a settlement statement record exists in the material. Observe whether a deduction conclusion is produced without statement evidence.
- Treating a similarity score as a permission filter. Input: a score value returned by retrieval. Compare that score with whether the material subject belongs to the same order. Observe whether a high score alone is taken to mean the record is usable. Similarity does not carry entity isolation.
How to verify the setup
- For the return record text and the draft sentence, input both and compare "pending review" with "return has been completed." The condition is met when the draft does not write pending review as completed and clearly marks the status as needing verification.
- For the fee-deduction draft sentence, input the original text and the draft. Compare whether a settlement statement field exists. The condition is met when the output states the settlement statement was not provided and the fee deduction needs verification.
- For the handover request, input that request with the material above. Compare the request's "confirm the fee has been deducted" with the material's evidence scope. The condition is met when the output separates facts already present in the material from business actions that require downstream verification. A paper input of [] can be used to check downstream handling of an empty result; it does not mean that a fictional rule retrieval will return empty.