What the documents look like
Reverse returned-goods inspection handles returned parcels and product grading rather than new outbound orders. Its working materials come from the WMS returns module, platform after-sales cases, carrier return records and the inspection bench. The daily documents are the return warehouse receipt, inspection photographs, the grade decision table and the resale listing sheet. Data changes per item and per day.
For the contract and transaction material comparison task, the return warehouse receipt, the after-sales work order, the carrier return record and the inspection conclusion must be checked under the same subject, the same period and the same field. Citation tracing requires every conclusion to point back to a specific document and field. A retrieval hit only supplies candidate source text. It does not constitute authorization or a business determination. A retrieved match shows textual similarity, while whether a contract took effect, whether a refund was approved and whether a return was completed all have to be checked against the original text, with missing evidence kept as missing.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model, and it outputs a citation array whose length may be zero. A zero-length array does not break the downstream chain. The citation limit is measured in tokens. The minimum relevance value applies only during semantic retrieval or when result reranking is used. A similarity score is a relevance measure and cannot serve as a filter for entity, permission or business ownership.
Hybrid retrieval merges multiple result paths through RRF. Reranking needs a model and is enabled according to settings. Image retrieval needs image input and the corresponding model capability. The optional paths listed in the overall retrieval flow are not all executed on every request.
Hypothetical paper sample for checking. The source text to be compared consists of three sentences. First: "Return warehouse receipt RT-2026-0417, subject is after-sales warehouse A, item is a certain model of electric kettle, returned quantity 1, status is signed for and awaiting inspection." Second: "Platform after-sales work order SH-2026-0417-02, same subject and same item, after-sales status is under review." Third: "Carrier return record CR-2026-0417, same waybill number, signed date 17 April 2026." The draft sentence to be judged is: "This return has completed inspection and the refund has been approved." This sample is a paper exercise and does not refer to a real customer or a real document number.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Comparison method and reason |
|---|---|---|
| Associated knowledge bases | Select only knowledge bases using the same vector model that contain the return receipt, after-sales order and carrier record | Check the vector model identifier first; exclude bases with a different model and do not run cross-model retrieval to compare missed recalls |
| Search mode | Use full-text or hybrid retrieval for numbered fields, semantic retrieval for descriptive fields | Numbers, models and waybill numbers suit keyword hits; natural-language descriptions suit semantic recall |
| Citation limit | Set a hypothetical starting value covering the tokens of the three sample source sentences | The starting value is estimated from the sample text length and is compared with the actual cited content length returned before adjustment |
| Minimum relevance | Set a hypothetical middle value, effective only during semantic retrieval or reranking | Compare with the number of returned citation entries; too low admits irrelevant fragments, too high may drop candidates |
| Result reranking | Enable the rerank model according to settings when candidates are many and the question is complete | Compare with ordering before reranking; the rerank score is used for filtering, not for permission judgment |
| Citation handoff | Check whether the search node citation array is passed to the answering step | Compare each answer sentence against the cited source text; check a zero-length citation chain against the downstream design |
Three common errors
- Treating a retrieval hit as authorization or proof of effect. Input is the three source sentences above plus the draft sentence. Compare the after-sales order status "under review" with the draft phrase "the refund has been approved." The retrieval hit only shows the text was recalled. The after-sales status has not changed to approved, so the draft sentence exceeds what the source supports.
- Treating a similarity score as an entity or permission filter. Input is a retrieval result from a knowledge base with the same vector model. Compare the returned entries with the target document number. A high or low score does not distinguish subjects and does not exclude records without access rights. Document number, subject and period require manual checking.
- Merging pending review, signed-for and inspected into one fact. Input is the return receipt status "signed for and awaiting inspection" and the after-sales order status "under review." Compare the values in the same field. Signing, inspection and refund approval are different stages. When the inspection completion record and the refund approval record are absent, the draft sentence is not supported.
How to verify the setup
- Input the three sample source sentences and the draft sentence. Compare "has completed inspection and the refund has been approved" in the draft with "signed for and awaiting inspection" and "under review" in the source. The expected observation is that the draft sentence is not supported by the source, both inspection completion and refund approval lack evidence, and the item is recorded as pending verification.
- Input a retrieval request against the knowledge base with the same vector model. Compare the returned citation array with the three sample documents. The expected observation is that each citation entry points back to the source text, that a zero-length citation array still lets the downstream chain run, and that no business conclusion is drawn from this.
- Input two knowledge base identifiers that use different vector models. Compare whether the vector model identifiers match. The expected observation is that the base with a different identifier is excluded, no cross-model retrieval is run, and only the identifier check result and the exclusion reason are recorded.