What the documents look like
Reverse returned-goods inspection handles returned parcels and item grades. It does not handle new-order outbound picking and packing, which follows a separate data path. This page addresses order fulfilment and after-sales question answering within the multi-turn prompts step. The working materials are the returned-goods inbound receipt, inspection photographs, the grade determination table, and the re-sale listing sheet. Data sources include the WMS returns module, platform after-sales records, carrier return parcels, and the inspection station. Updates occur per item and per day. These materials set the boundary of what a multi-turn answer may state: the inbound receipt records the returning party and the inbound scope, the photographs and grade table record the basis for judging item condition, and the re-sale listing sheet records downstream disposition. Answers must be organised around these materials, and business facts outside the material scope remain unknown.
A returned-goods inbound receipt carries a status field. When that field reads pending inspection, no grade conclusion exists yet. A grade determination table may require a numeric measurement before a grade can be assigned. An inspection photograph may record a defect without recording the measurement the table requires. Where a required value is absent, the determination condition is unmet, and a downstream re-sale listing request has no supporting basis.
Constraints at the multi-turn prompts step
Multi-turn dialogue depends on four elements working together: history records, question optimisation, model and length settings, and reply format. History records determine how many prior turns are carried into context, and a new session differs from a continued session in what is supplied. Question optimisation handles reference resolution and question expansion in continuous knowledge-base dialogue, and short questions do not all benefit from it. Model and length settings determine the referenceable context and the reply ceiling. Reply format is chosen according to what downstream processing requires; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure.
The prompt must state the material scope, the answering task, and how to handle missing evidence, so that the model does not supply business facts on its own. Retrieval results provide reference content only. They carry no permission or entity isolation, and a similarity setting does not exclude unrelated businesses or unauthorised records. Retrieval is not authorisation.
Hypothetical drill sample: the returned-goods inbound receipt RH-2026-0417-03 records a platform buyer as the returning party, a vacuum flask as the item, a quantity of 2, an inbound date of 17 April 2026, and an inspection status of pending inspection. The grade determination table states that a surface scratch with a depth below 0.5 mm may be graded as good, and a depth at or above 0.5 mm is graded as defective. An inspection photograph records one scratch on the flask body and does not annotate depth. The draft sentence awaiting review reads: "This return has been graded as good and a re-sale listing sheet can be generated directly." The sentence lacks the depth value, the grade table requires a depth number, and the photograph does not annotate depth, so the determination condition is unmet.
Configuration and implementation checks
The table below lists settings already described in the official documentation, or implementation checks, tied to the sample above. The values are starting points for this hypothetical trial, used for comparison against the actual values in the sample. They are not official defaults.
| Configuration item | Hypothetical trial starting point | How to check |
|---|---|---|
| Memory turns | Keep a small number of history turns | Compare the context carried into a continued session against the input of a new session |
| Reply ceiling | Set a starting point matching the sample answer length | Compare against the sample draft sentence length and observe whether truncation occurs |
| Temperature | Take a low value | Compare against stable customer-service and knowledge-base scenarios |
| Reply format | Keep the default | Change it when downstream processing reads a fixed format, then check the returned structure |
| Multimodal recognition | Enable image recognition according to model capability | Check whether the inspection photograph is read as image input |
| Thinking configuration | Take the default | Compare against the sample question complexity; do not treat the thinking budget as a citation limit |
Three common errors
- Treating a pending-inspection status as a completed determination. In the sample, the inbound receipt inspection status is pending inspection, while the draft sentence states that the return has been graded as good. Both values sit on the same subject and the same field and are mutually exclusive, so the draft sentence is unsupported.
- Substituting the photograph for the grade determination table. The photograph does not annotate scratch depth, the grade table requires a depth number, and without that number the determination condition is unmet, so no re-sale listing sheet can be generated.
- Treating a retrieval result as authorisation. Retrieval returns reference content and does not indicate that the return has been approved for re-sale. Generating a re-sale listing sheet requires its own downstream implementation.
How to verify the setup
- Input the sample returned-goods inbound receipt text and the draft sentence. Compare the inspection status field value "pending inspection" against the draft sentence statement "graded as good". If the two values sit on the same subject and the same field and are mutually exclusive, the draft sentence does not hold and should be changed to pending verification.
- Input the sample grade determination table and the inspection photograph. Compare the depth number required by the table against what the photograph actually records. If the photograph does not annotate depth, the determination condition is unmet and no re-sale listing sheet should be generated.
- Input the sample re-sale listing request. Compare the determination conclusion the request relies on against the existing materials. If the materials support only a pending-inspection status, the request remains pending verification and requires the depth data to be supplied before any judgement is made.