What the documents look like
Trade returns and claims handle reverse logistics and liability determination after export. The working materials come from customer email, quality inspection reports, customs systems, and insurance claim systems. The day-to-day documents are the return agreement, the claim letter, the quality report, and the return customs declaration. These materials are produced case by case on an irregular schedule, and the same shipment may carry both Chinese and English versions alongside several rounds of revision.
Citation tracing answers one question: which passage of which material does a statement in the answer come from, and do that material's version and scope of application support the conclusion being drawn. The retrieval side does not create batch synchronization or real-time triggers from the case-by-case update pattern. A retrieved passage is not an approved passage, a signed passage, or an effective passage.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero; the output path still completes when nothing is found. Hybrid retrieval merges multiple result paths through RRF. Result reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and the corresponding model capability. The minimum relevance value is a filter threshold, not a final ordering rule, and it does not perform permission or entity isolation. The citation limit is measured in tokens and is unrelated to any thinking budget.
Hypothetical drill sample: in one return case, the material text supplied for checking reads: "Return agreement (draft) states: the buyer must raise a written quality objection within 15 days of arrival; this agreement takes effect on the date both parties sign it. Quality inspection report states: 3 cartons sampled, 1 carton shows appearance defects." The draft sentence to be judged reads: "The return agreement is in effect and the buyer's claim request has been supported." When comparing, note that draft text, signature status, whether a claim was submitted, and whether it was supported are separate facts. The source text does not record a signing date or a claim submission record.
Configuration and implementation checks
| Configuration or check item | Use on this page | Correspondence with the sample |
|---|---|---|
| Linked knowledge bases | Select knowledge bases with the same vector model as the retrieval scope | The return agreement and quality report must sit in the same searchable library |
| Search mode | Semantic search, full-text search, or hybrid search | Number and date fields suit full-text search |
| Minimum relevance | Filters low-relevance results; applies only in semantic search or with reranking | Used to screen out passages unrelated to the return case |
| Citation limit | Maximum tokens cited per search | Controls the length of citation content returned to the model |
| Question optimization | Coreference removal and question expansion before retrieval | Completes what "that agreement" refers to in multi-turn follow-ups |
| Citation content output | Outputs citations as an array, length may be zero | An empty array is used to check the downstream handling design |
Hypothetical trial starting point for this page: set minimum relevance to 0.5, on the reasoning that the sample question carries a clear subject and date field and needs comparison against a lower value to observe changes in recall range. This value is not an official default.
Three common errors
- Treating a retrieval hit as evidence of authorization or effectiveness. Input is the draft text above and the draft sentence to be judged; compare whether "draft" and "in effect" belong to the same field and the same state; the observed result is that the source text records no signing date, so the judgment stays pending verification.
- Treating an empty citation array as a workflow error. Input is a paper query that returns no hit; compare the output against the documented statement that the length may be zero; the observed result is that the path still completes, and the downstream design for an empty array needs checking.
- Forcing knowledge bases with different vector models into one linkage. Input is two libraries with different vector model identifiers; compare the model identifiers; the observed result is that they do not meet the shared-linkage condition, so they are excluded and recorded, and no cross-model retrieval comparison is run.
How to verify the setup
- Citation source check. Input is the draft return agreement text and the quality report text above; compare "in effect" and "has been supported" in the draft sentence against the state fields recorded in the source text; the condition is met when the draft sentence matches the source text or is explicitly marked as pending verification.
- Empty citation handling check. Input is an empty array returned by a paper query; compare whether the downstream node handles a citation array of length zero; the condition is met when the downstream does not stop on the empty array and does not fabricate citation content.
- Citation display check. Input is one retrieval that hits the quality report; compare the text cited in the answer against the quality report source text; the condition is met when each cited passage can be traced back to its source paragraph, without substituting answer fluency for source evidence.