What the documents look like
Returned-goods inspection handles reverse parcels and product grading. It does not handle outbound picking or packing for new orders. The working materials are the return warehouse receipt, inspection photographs, the grade determination table, and the resale listing sheet. Data arrives from the WMS returns module, platform after-sales records, carrier return shipments, and the inspection bench, with updates per item and per day.
For copy drafting, the condition, functional state, and accessory status quoted in a draft sentence must agree with the inspection photographs and the grade determination table. Those two materials are the factual source for the draft; the retrieval step brings the corresponding material into the drafting flow. This page addresses one task only: drafting product and marketing copy for returned-goods inspection, at the knowledge-base retrieval and recall step. Whether a specific promotional sentence is supported by the materials is the question here, and it stays separate from identity verification, report analysis, or any other task.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. Check the model identifier first, exclude any base that does not match, and record what was actually observed. Retrieval output is an array of references whose length can be 0; the output path still runs even when nothing is found.
The reference cap is measured in tokens. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. Semantic recall and relevance scores do not carry authorization or entity isolation, and retrieval is not authorization. A relevance score is a filtering threshold, not a category, entity, or permission filter.
Hypothetical paper exercise sample. The source text reads: "Return warehouse receipt RT-2026-0417 records: product A, return reason 'customer seven-day no-reason return', signed-for date 10 April 2026. Inspection photograph IMG-0417-03 shows a scratch of about 2 cm on the upper-left corner of product A's casing, and the screen lights up normally. Grade determination table GR-2026-0417-02 grades product A as grade B, with the note 'minor cosmetic scratch, function normal'." The draft sentence to be checked is: "Product A is brand new and unopened, fully functional, with no scratches on the casing." The comparison object is the draft sentence against those three materials. The grade table records grade B and the photograph shows a scratch, while the draft sentence states brand new, unopened, and no scratches. The same subject, same field carries an explicit, mutually exclusive value, so the promotional sentence is not supported by the materials.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the return warehouse receipt base, the inspection photograph description base, and the grade determination table base | All three use the same vector model; check each vector model identifier and exclude any that do not match |
| Search mode | Hybrid retrieval | Copy checking involves both natural-language description and identifiers such as receipt numbers and grade codes; compare recalled content against pure semantic retrieval |
| Reference cap | Set to a token amount covering the full text of one grade determination table | This is a hypothetical starting point for this exercise; test with materials of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Take a value after labelling relevant and irrelevant samples first | Applies only under semantic retrieval or result reranking; compare whether irrelevant fragments are filtered out at different values |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether subject or receipt number was altered |
| Result reranking | Enable per settings when a rerank model is available | Compare candidate order before and after reranking; the rerank score range is 0-1 |
Three common errors
- Treating grade B in the grade determination table as brand new. The error is ignoring a mutually exclusive value in the same field: the grade table records grade B with a note of minor cosmetic scratch, while the draft sentence says brand new and unopened. The check inputs the grade determination table text and the draft sentence, compares the grade field and the cosmetic description field, and if the two values are mutually exclusive the draft sentence is not supported.
- Using a retrieval score in place of material checking. The error is treating a similarity score as factual grounds. The check inputs the draft sentence and the recalled reference content, compares whether the reference contains subject, object, and state, and if the reference does not contain the grade field, records an evidence gap rather than a negative finding.
- Writing a pending review state as already approved. The error is conflating distinct facts. The check inputs the handover request and the review record, compares the request state against the approval record, and if only a pending request exists, records it as pending verification rather than approved.
How to verify the setup
- Input the three hypothetical source sentences and the draft sentence "Product A is brand new and unopened, fully functional, with no scratches on the casing." Compare the draft sentence against the grade field of GR-2026-0417-02 and the photograph description. If the grade is B and the photograph shows a scratch, the draft sentence is not supported and must be returned for revision.
- Input an explicit paper empty result []. Check whether the downstream drafting flow still runs, and compare whether the output path is interrupted by the empty reference. If the path runs through, this matches the documented behavior for empty references.
- Input a query containing receipt number RT-2026-0417. Compare the recalled content under hybrid retrieval against pure semantic retrieval. If the exact identifier query is matched under hybrid retrieval and not under pure semantic retrieval, record that difference. All of the above are paper checks; no test was actually executed.
Cross-model bases cannot be associated together. Check the identifier, exclude non-matching bases, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls. The optional paths in the overall flow diagram do not mean every request executes all of them: image retrieval requires a corresponding image input and model capability, and reranking requires an available model enabled per settings. RRF is a fusion process for mixed results, not a toggle switch. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.