What the documents look like
Trade returns and claims work runs in the reverse direction from ordinary export declaration. The materials are generated after goods have left, and they concern reverse logistics and the allocation of responsibility. A single case typically brings together a return agreement, a claim letter, a quality report and a return customs declaration. These arrive from separate places: customer email, inspection reports, customs systems and insurer claim systems. Updates are irregular and case by case, so the knowledge base is not refreshed on a fixed rhythm.
For the comparison task, the practical difficulty is that the documents within one case are often inconsistent with each other. The parties named, the goods batch and the time reference can differ from one file to the next. Retrieval does not resolve those differences. The knowledge base search step returns candidate passages, and a person then checks the fields against each other. Retrieval does not decide which side is liable, and it does not file a return declaration or submit an insurance claim.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. This condition applies to vector search itself. It does not extend to customs systems, business registration systems or insurer systems, which are separate from the knowledge base.
Cited content is returned as an array, and the length can be zero. When nothing is found, the output branch still runs. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple recall paths through RRF, and ReRank then performs a second ordering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation.
Field comparison across the return agreement, claim letter, quality report and return customs declaration is a manual step. Retrieval supplies candidate source passages only.
Hypothetical paper exercise. The source passages to be compared are as follows. Return agreement RT-2024-018: seller Hongyuan Machinery, buyer Nordic Equipment Company, returned goods hydraulic pumps, quantity 12 units, return reason bearing noise, agreement signing date 5 March 2024. Quality report QR-2024-033: same batch of hydraulic pumps, quantity 12 units, inspection conclusion bearing noise, report date 2 March 2024. Claim letter CL-2024-007: claimant Nordic Equipment Company, respondent Hongyuan Machinery, claim amount USD 8,600, claim basis quality report QR-2024-033, letter date 8 March 2024, claim letter status pending review. The draft sentence to be assessed is: "Return agreement RT-2024-018 has confirmed quality responsibility and completed the claim." This draft sentence involves three checks of party, object and status. The parties are Hongyuan Machinery and Nordic Equipment Company. The object is hydraulic pumps, 12 units. The status is that the claim letter is pending review.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link only the return and claim materials base already confirmed to use the same vector model | Check the vector model identifier first, exclude bases with a different model, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare with semantic retrieval alone and full-text retrieval alone, and observe recall differences on identifier-type fields |
| Citation limit | Set to 3000 tokens as the hypothetical starting point for this exercise | Compare with the actual length of the return agreement plus the quality report, and observe whether citations are truncated |
| Minimum relevance | Set to 0.4 as the hypothetical starting point for this exercise | Compare with scores of annotated relevant and irrelevant samples, and observe the filtering effect |
| Question optimization | Enabled | Compare with the disabled state, and observe whether the subject of a question containing a reference is altered incorrectly |
| Result reranking | Enabled according to the setting | Compare with results without reranking, and observe ordering changes when candidates are numerous |
The citation limit and minimum relevance values are starting points for this hypothetical exercise, not official default values. They require comparison with actual material length and annotated samples before adjustment.
Three common errors
- Treating retrieval recall as a determination of responsibility. The input is a return agreement passage returned by retrieval. The comparison object is the responsibility wording in the claim letter. The observable result is whether the two name the same responsible party. If the claim letter is only pending review, a retrieval hit does not constitute a confirmation of responsibility.
- Linking knowledge bases that use different vector models. The input is the vector model identifier of two knowledge bases. The comparison object is the same-vector-model condition required by the node. The observable result is whether the identifiers match. When they do not match, exclude that base and record the actual observation without inferring a runtime outcome.
- Treating a pending-review status as a completed claim. The input is the status field of claim letter CL-2024-007. The comparison object is the wording "completed the claim" in the draft sentence. The observable result is whether the status field states completion. Pending review does not show that the claim has been executed.
How to verify the setup
- Input the return agreement passage from the hypothetical sample and the draft sentence. Compare the party, object and status in the draft sentence against the source passage for the same party, the same period and the same field. The observable result is whether the draft sentence turns a pending-review status into a completed one. If the source does not state a confirmation of responsibility, record it as unsupported.
- Input quality report QR-2024-033 and claim letter CL-2024-007. Compare the report number reference and the date reference in both. The observable result is whether the number matches and whether the dates fall within the same period. Record missing fields as evidence gaps.
- Input an explicit paper empty result [] and compare how the downstream node is designed to receive that output. The observable result is whether the output branch runs. When retrieval returns nothing, do not assert that a fictional rule necessarily returns []; check only the downstream handling of an empty array.
The knowledge base search node, its same-vector-model condition, the array output with length zero, the search modes, RRF merging, ReRank ordering, the citation limit in tokens and the minimum relevance scope are documented behaviours. The citation limit value, the minimum relevance value, the search mode choice, question optimization and result reranking are settings or implementation checks for this exercise. Field-by-field comparison across the four document types is a manual review action. Filing a return declaration, submitting an insurance claim and confirming liability are business actions that require separate downstream implementation, and no such action is claimed here.