What the documents look like
Returned-goods inspection handles reverse parcels and item grading. It does not handle outbound picking for new orders. The working materials come from the WMS returns module, platform after-sales records, carrier return parcels and the inspection bench. The daily documents are the return inbound note, inspection photos, the grade decision table and the resale listing note. Store service and internal training cite these materials, so each cited item is checked against its source document.
Three properties of this category shape the citation task. First, the unit of work is one returned item, so a citation that covers a batch can hide which item the grade belongs to. Second, the grade decision is a judgement recorded on a specific date by a specific bench, so a later draft sentence that drops the date or the bench loses the traceable link. Third, resale or scrap decisions sit downstream of inspection, so a citation that shows a grade does not by itself show that resale was authorised. The brief lists daily and per-item update frequency; nothing in this page depends on a batch, real-time or transaction-triggered mechanism.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The citation output is an array and its length may be zero, which means the downstream chain still runs even when nothing was retrieved. The minimum relevance value is a filter threshold and applies only in semantic retrieval or when result reranking is used; it is not a permission or entity isolation control. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires image input and a model capable of handling it. The optional paths drawn in the overall retrieval flow diagram are not all executed on every request.
A relevance score does not separate unrelated businesses or unauthorised records, and retrieval is not authorisation. A pending review request is not an approved request. An unsigned document is not a completed document.
Hypothetical paper exercise sample. A return inbound note for an after-sales warehouse (note number HT-2024-0312, dated 12 March 2024) states: "Item A outer packaging damaged, inner contents complete, graded B resalable." Inspection photos from the same batch are annotated: "photographed 2024-03-12, Item A front surface scratch." The grade decision table records: "Grade B: appearance defect, function normal." A draft sentence awaiting review reads: "Item A has been confirmed by inspection to function normally and can be placed directly on resale." This request has not been authorised for execution and remains at the pending review stage.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select the return inbound note base and the grade decision table base | Both bases must use the same vector model; compare their model identifiers and exclude any base whose model differs |
| Search mode | Hybrid retrieval | Note-number style queries need semantic and full-text paths together; compare recall coverage against semantic retrieval alone |
| Minimum relevance | Set to 0.5 | Hypothetical starting point for this exercise, not an official default; compare the number of items remaining after filtering against 0.3 |
| Citation limit | Set to 2000 tokens | Hypothetical starting point; compare citation completeness against 1000 tokens |
| Question optimisation | Left disabled | The exercise question is a single clear turn; compare elapsed time against the enabled state |
| Result reranking | Left disabled | The candidate set is small; compare ordering against the enabled state |
The citation limit uses the documented token unit. It is not the same as a thinking budget, and a thinking budget setting does not control citation length. The optional image retrieval path is not exercised here because the sample contains no image input to the model, and image retrieval would require both an image input and a model able to process it.
Three common errors
- Treating the minimum relevance value as a permission filter. Input two knowledge base identifiers that use different vector models and compare whether the system allows them to be associated; observe whether association is refused. Even when association is allowed for same-model bases, a relevance threshold does not exclude unauthorised records, and no run result should be asserted for the cross-model case beyond the identifier check itself.
- Treating an empty citation array as a failure. Input a paper-level empty array
[]and compare it against the downstream node configuration; observe whether the chain continues to execute. The documented behaviour is that a zero-length citation output still lets the chain run, and no error should be assumed. - Treating a pending review request as authorised. Input the draft sentence and the return inbound note text, compare the note field "graded B resalable" against the draft phrase "can be placed directly on resale", and observe whether an authorisation signature record exists. The draft is at pending review, so the absence of a signature is a retained evidence gap rather than proof that permission is missing.
How to verify the setup
- Input the return inbound note text and the inspection photo annotation, then compare "outer packaging damaged, inner contents complete" against the draft phrase "function normally". Observe whether the draft phrase is supported by the source text. The note does not record a functional test, so the function claim remains unsupported and pending verification.
- Input the grade decision table entry "Grade B: appearance defect, function normal" and the draft sentence, then compare the grade field against the resale conclusion. Observe whether an authorisation boundary statement is present. When it is absent, record the evidence gap rather than supplying an authorisation clause.
- Input the paper-level empty array
[]together with the downstream node configuration, then compare whether the chain proceeds when the citation length is zero and observe whether any error appears. Do not assume that a retrieval rule must return an empty array; the empty input is a paper exercise of the downstream design only.
Across all three checks, the comparison must use the same subject, the same period and the same field. A field with two mutually exclusive values is a contradiction; a field that is simply not recorded, or that only carries evidence from a related stage, is unsupported or pending verification. A quotation validity period does not state a supply period, and reaching agreement, signing and contract effect are three separate facts. When the material needed for a check is not inside an accessible retrieval scope or a clearly defined manual input, record the gap and leave it open.