What the documents look like
Brand after-sales repair work is organized around warranty responsibility and physical repair points, which separates it from platform after-sales handling that processes refunds. The working materials for this category are warranty cards, repair work orders, inspection reports, and exchange approvals. These records come from after-sales systems, store orders, repair-point feedback, and the national three-guarantees catalog. Together they determine what state a device is in: whether it sits inside a valid warranty period, whether an inspection conclusion supports part replacement, and whether an exchange approval has been signed off. The multi-turn prompt step has to state these states accurately and keep the model on the same subject, the same work order, and the same field across follow-up turns.
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. The model must not supply a warranty conclusion, an approval result, or a repair-point commitment that the materials do not contain. For history, the actual context carried into a turn has to be checked, and a new session has to be distinguished from a continued session. Problem optimization is checked only in knowledge-base continuous-conversation settings for reference resolution and question expansion, and short questions are not assumed to behave the same way everywhere. For model and length, the supported context and reply settings are checked against long inspection reports combined with follow-up questions. Reply format follows what the downstream step needs; when json_schema is chosen, the format requirement is filled in and the returned structure is verified.
Hypothetical paper exercise: a brand after-sales center holds work order W-2024-0912. The customer submits a warranty card copy showing purchase date 10 March 2023 and a whole-machine warranty period of 12 months. The inspection report states "mainboard failure, not caused by human action." The exchange approval field is blank. A proposed sentence reads: "Upon review, the device is within the warranty period and the damage was not caused by human action; exchange approval has passed, so a replacement can be arranged." The pending request is: "Please reply to the customer about the replacement time based on the above materials."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model that supports long context | Inspection reports and work orders are long; compare whether the model's maximum context covers the material length |
| Memory turns | Set to 3 turns | Customer-service settings usually keep a small history; compare with the default value to see whether it is enough |
| Temperature | Set low | Warranty judgment needs stability; compare with a higher value to see whether the answer drifts |
| Reply cap | Enabled and raised as needed | A fuller explanation is needed; compare against truncation behavior |
| Reply format | Keep default | No downstream step reads a fixed structure; compare with json_schema to see whether it is necessary |
| Stop sequence | Not set | Ordinary question answering does not need a fixed end marker |
Three common errors
- Reading a blank exchange approval field as "approval passed." Input is the blank approval field on work order W-2024-0912; the comparison is between "passed" and "not stated"; the observable result is that the proposed sentence contains an approval conclusion the source does not carry.
- Treating "not caused by human action" as equivalent to "within the warranty scope." Input is the inspection report phrase "not caused by human action"; the comparison covers the warranty period on the warranty card and the nature of the damage as two separate fields; the observable result is an answer that merges two fields into one conclusion.
- Carrying a new work order's question into an old work order's context in a continued session. Input is two turns under the same chatId with different work order numbers; the comparison is between the subject identifiers of the two turns; the observable result is an answer that cites the previous work order's inspection conclusion.
How to verify the setup
- Take the exercise source text and input the proposed sentence "exchange approval has passed." Compare it against the approval field state in the source. The sentence is supported only when the approval field shows an explicit signed record; here it is blank, so record it as unsupported.
- Take the exercise source text and input the follow-up "is this device within the warranty period." Compare the purchase date and warranty period fields on the warranty card. Answer in-warranty only when both fields jointly support it; the inspection report does not constitute warranty-period evidence.
- Take the work order number from the exercise and a second work order number, and input two consecutive turns under the same chatId. Compare the subject identifiers of the two turns. Carry context forward only when the subject matches; when the subject differs, record it as needing reconfirmation and do not perform a cross-work-order reference.