What the documents look like
For after-sales questions in Seller ratings management, this page examines FastGPT multi-turn prompts. Separate complaint facts, platform metrics and correction receipts across follow-ups. The examples are hypothetical.
Constraints at the multi-turn prompts step
Suppose a complaint note records contact with the buyer, a score extract identifies its reporting period, and a correction log awaits a reply. A progress question followed by a recovery question must separate individual contact, period metrics and correction evidence. Earlier contact wording establishes neither buyer acceptance nor improved platform scores. This exercise supplies no actual score, change magnitude or platform review finding. It tests preservation of state and reporting period in invented text.
Configuration and implementation checks
| Review-maintenance turns | Hypothetical record | Boundary |
|---|---|---|
| Ticket | Contact made | No acceptance or withdrawal inference |
| Metric | Period extract | Do not replace with question date |
| Correction | Reply pending | No acceptance claim |
| History summary | May omit conditions | Return to original log |
| Outcome | No actual score supplied | Invent no recovery magnitude |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Prompt and task | Define sources, the task and evidence-gap handling. |
Conversation history | Inspect the history actually submitted. |
Question optimization | Review reference resolution and question expansion. |
Model and length | Inspect context and reply length separately. |
Reply format | json_schema needs its format and parsing validation. |
Three common errors
- If the reporting period disappears and the answer claims recovery this week, compare the source period with actual input. The question date cannot replace the data period.
- If contacting the buyer becomes withdrawal of the review, inspect the contact note, buyer response and review state. Contact alone creates no outcome.
- If an intermediate history summary converts a pending correction reply into acceptance, compare the source log and receipt with that summary’s omitted condition.
How to verify the setup
Ask first what was done, then what evidence is missing, and finally add a proposed buyer response. Each turn should change only with the corresponding new evidence. Repeat in a fresh conversation to expose dependence on earlier summaries. Instructions should answer ticket state, review state and period metrics separately. Actual recovery needs matching platform records. This paper exercise changes no review, withdraws no complaint and sends no buyer message; drafted wording establishes no performed contact.
| Material context | Scope |
|---|---|
| Document types | Negative Review Handling Records, Experience Score Details, Service Diagnosis Report and Rectification Ledger |
| Candidate sources | Douyin Shop Experience Score, Tmall Service Metrics, JD After-Sales Evaluation and Customer Service System |
| Technical baseline | FastGPT v4.17.2 |