What the documents look like
Returned-goods inspection handles parcels that were sold, sent back, and received into a reverse warehouse. The work object is the physical return together with its accompanying paperwork, and it does not cover outbound picking for new orders. The main data sources are the WMS return module, platform after-sales records, carrier return information, and the inspection station. The documents in scope are the return inbound note, inspection photographs, the grade determination sheet, and the resale listing sheet, updated per item and per day.
For copy drafting, the grade description, defect statement, and sellable status in a proposed sentence must trace back to specific fields in these documents. Inspection photographs and the grade determination sheet usually record appearance evidence and conclusion separately, and the two may not both be complete for the same return, so citation requires item-by-item correspondence.
Constraints at the citation tracing step
The drafting step takes retrieved document content as citation sources and forms proposed sentences from them. The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs citation content as an array whose length may be 0. When nothing is retrieved, that output path still runs through. The citation limit counts tokens, and the minimum relevance value applies only under semantic retrieval or result reranking. Hybrid retrieval merges multiple result paths through RRF, and reranking requires an available model enabled by configuration. Retrieval itself is not authorization, and a similarity score does not perform entity or permission filtering. The return inbound note and the resale listing sheet may sit in different knowledge bases, and they can be associated together only when the vector model identifier matches. When the identifiers differ, exclude them first and do not run cross-model retrieval to compare missed recalls.
Hypothetical drill sample: a return inbound note states "Return ID RT-0001, item is a vacuum flask, appearance status is scratches on the body, inspection conclusion is downgraded sellable, date 5 March 2024." The inspection photograph note for the same batch states "the scratch is on the front of the body, roughly two centimeters long." A proposed sentence awaiting judgment reads "This vacuum flask has no scratches on the body and is listed as new." The check compares the proposed sentence against "scratches on the body" in the return inbound note and the scratch location description in the photograph note, requiring the subject, object, status, and date range to agree.
Configuration and implementation checks
| Configuration or check item | Purpose on this page | Basis and boundary |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases holding the return inbound note, inspection photographs, grade determination sheet, and resale listing sheet | Only knowledge bases with the same vector model can be associated together; check the identifier first |
| Search mode | Use full-text retrieval for document numbers or model numbers, semantic or hybrid retrieval for natural-language status descriptions | Hybrid retrieval merges through RRF; reranking requires a model enabled by configuration |
| Citation limit | Control the maximum tokens of content cited per search | Counted in tokens; a hypothetical starting value for this drill is set close to the body length of one return inbound note, so truncation can be compared |
| Minimum relevance | Filter low-relevance results | Applies only under semantic retrieval or result reranking, and does not act as a permission filter |
| Citation content output | Carry retrieval results and support item-by-item comparison against source text | The array length may be 0; an empty array does not mean the workflow errored out |
| Reply format | Select json_schema and fill in the format requirement when a fixed structure is needed | Change this only when a downstream step must read a fixed format |
Three common errors
- Treating a status description in the proposed sentence as already supported by the materials. Input is the proposed sentence "no scratches on the body" together with "scratches on the body" in the return inbound note. Compare the value of the same field; the two values are mutually exclusive, so the sentence contradicts the material and must be sent back for revision.
- Using a similarity score to infer permission or entity scope. Input is the citation array and scores returned by one search. Compare the score against the authorization field; the score does not perform permission filtering, authorization needs a separate system implementation, and this item is recorded as pending verification.
- Treating an empty citation array as a workflow failure. Input is citation content output
[]. Compare whether the downstream node is still triggered; the path still runs through with an empty array, so check whether the downstream step is designed for empty input.
How to verify the setup
- Take the return inbound note text and the proposed sentence from the hypothetical sample, run retrieval, and compare the citation content. Check whether the citation includes "scratches on the body" and the date 5 March 2024. If the citation is missing that field, judge the sentence unsupported and record the evidence gap.
- Take the inspection photograph note and the proposed sentence, and check whether the citation includes the scratch location description. If retrieval does not return that entry, record it as not recalled, and do not infer that the material does not exist.
- Use an explicit paper input of
[]to check the downstream design and observe whether the flow continues. If it continues, empty-result handling is in place; do not claim that a fictional rule search must return an empty array.
The public technical facts checked for this feature are: the knowledge base search scope should be verified against the selected knowledge bases and material versions, and only knowledge bases with the same vector model can be associated together; the search node outputs a citation array whose length may be zero, and an empty array does not mean the workflow has errored out; check whether retrieval output is handed to the answering step and compare answers against citation source text item by item; use traceable samples to check source display, and do not substitute thinking process or answer fluency for source evidence; test sources with both current and historical document versions, and whether a citation is adopted still requires review against the effective version.