What the documents look like
Returned-goods inspections handle reverse parcels and item grades. The work does not pick or pack new outbound orders. Data comes from the WMS returns module, platform after-sales records, carrier returns and the inspection bench. The documents in daily use are the return inbound note, inspection photographs, the grade decision table and the resale listing note. Records update per item and per day.
These materials carry different roles during operations handover. The return inbound note records the returning party, the item object and the quantity. Inspection photographs and the grade decision table record the judged condition of the goods. The resale listing note records the intended downstream disposition. Retrieval brings these materials to the handover recipient. A retrieval result supplies content only. It does not grant any authority, and it does not confirm that a disposition has been approved or executed.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Search mode is compared across semantic retrieval, full-text retrieval and hybrid retrieval for a given test question. No single mode is assumed to suit all materials. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and the corresponding model capability. Citation content is output as an array, and its length can be zero; when nothing is retrieved, the output path still completes.
Hypothetical paper exercise sample. An after-sales warehouse inspector submits a return inbound note whose text reads: "Returning party: East China after-sales warehouse; returning object: batch R-0417 goods; status is pending inspection; quantity: 12 items; registration date: April 17." Among the same batch of materials, a grade decision table reads: "Batch R-0417, decided grade: B, decision maker: inspection bench A, decision date: April 17." A draft sentence reads: "Batch R-0417 has completed resale listing authorization and may be listed." A pending determination request reads: "Please confirm, based on the above materials, whether this batch has obtained resale listing authorization."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the single vector-model base holding the return inbound note and the grade decision table | First check the vector model identifier of each base, exclude bases with a different model, and do not run retrieval across models |
| Search mode | Run the same question separately under semantic retrieval and full-text retrieval | Compare whether the recalled content covers the two sample materials under each mode |
| Citation limit | Set a token amount large enough to hold the full text of the return inbound note and the grade decision table | Use the actual length of the two sample materials as the comparison basis and observe whether content is truncated |
| Minimum relevance | Start from a middle value, then observe relevant and unrelated sample materials separately | Compare the pass result of the two material groups only under semantic retrieval or result reranking |
| Question optimization | Enable for the unclear reference in "whether this batch has obtained resale listing authorization" | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Observe scores when a rerank model is available and enabled in the settings | Compare how rerank scores and vector scores order the sample materials |
Three common errors
- Treating a retrieval recall as confirmation of authorization. In the sample, the grade decision table records only a B grade decision, while the draft sentence states that resale listing authorization has been completed. The input is the grade decision table text and the draft sentence. The comparison is between the two different fields "decided grade" and "listing authorization." The observable result is that the authorization described in the draft sentence has no corresponding record in the materials, so it is unsupported.
- Forcing knowledge bases with different vector models into one association. The input is the vector model identifier of each candidate knowledge base. The comparison is whether the identifiers match. The observable result is that bases with non-matching identifiers are excluded, and the retrieval scope keeps only bases with the same model.
- Treating a pending-review status as a completed operation. In the sample, the return inbound note status is "pending inspection." The input is that status field. The comparison is between the values "pending inspection" and "authorization completed." The observable result is that the two values are mutually exclusive, and the authorization described in the draft sentence contradicts the registered status.
How to verify the setup
- Run retrieval against the sample draft sentence "Batch R-0417 has completed resale listing authorization and may be listed." The input is this draft sentence together with the return inbound note and grade decision table text. The comparison is whether the field "resale listing authorization" and its value appear in the materials. The observable result is that the materials state only "pending inspection" and "grade B," the authorization field is absent, and the finding is unsupported with the evidence gap preserved.
- Run retrieval against the pending determination request "Please confirm, based on the above materials, whether this batch has obtained resale listing authorization." The input is this request text. The comparison is whether the recalled content includes the return inbound note and the grade decision table. The observable result is that if both materials are recalled and the authorization field remains absent, then authorization cannot be determined from the available materials; if the recall is an empty array, record that no content was retrieved and do not infer an authorization status.
- Check cross-base association. The input is the vector model identifier of each candidate knowledge base. The comparison is whether the identifier matches the base holding the sample materials. The observable result is that non-matching bases are excluded, retrieval proceeds only within the same-model base, and the actually observed recall scope is recorded without comparing missed recalls from cross-model retrieval.