What the documents look like
For after-sales questions in Marketplace advertising, this page examines FastGPT multi-turn prompts. Separate order-attribution labels from fulfillment and service events while preserving the current order and source roles. The examples are hypothetical.
Constraints at the multi-turn prompts step
An attribution table’s conversion-channel label supports reading that field, without establishing shipment or completed after-sales handling. A proposed on-site search advertising label can test whether the prompt preserves this scope. If an answer converts successful conversion into completed service, trace its event source. Without that record, the service conclusion cannot be verified. The attribution field’s definition and period also need inspection; a channel label is not evidence for the entire customer journey.
Configuration and implementation checks
| Conversation or implementation check | Support provided by advertising sources | Acceptance trace |
|---|---|---|
| Attribution label | Channel field, separate from fulfillment events | No service-completion inference from conversion |
| Expected and actual | Separate plan from attribution | Retain period and field definition |
| Historical subject | Distinguish current and other orders | Match original question, history and rewrite |
| Service gap | Order F lacks a service record | Trace generated steps and status individually |
| Added source | Actual delivery of case or event material to the next turn | Record channel and service citations separately |
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 a conversion-channel field becomes dispatch or completed service, compare attribution wording and answer with order-event evidence. Conversion supplies no complete customer journey.
- If expected and attributed channels differ and one source is declared wrong, preserve definitions, periods and stages. Inspect whether optimization changed expectation into an event.
- If follow-up borrows another order or invents service steps without records, inspect original question, history and returned sources. Restore the current subject and keep missing tickets unresolved.
How to verify the setup
Order F has conversion evidence but no service record. A service-detail question needs the relevant case or event material. Since the model may still answer, inspect generated steps and completion states sentence by sentence. After adding the source, record whether it actually enters the next turn and repeat the original question and a follow-up. Retain channel interpretation separately from service evidence. Neither a continuing conversation nor a conversion label validates the complete customer experience.
| Material context | Scope |
|---|---|
| Document types | Advertising Plan, Keyword Report, Audience Package Description and Conversion Attribution Table |
| Candidate sources | Juliang Qianchuan, JD Kuaiche advertising, Alimama and Store Transaction Data |
| Technical baseline | FastGPT v4.17.2 |