What the documents look like
Trade returns and claims handle reverse logistics and responsibility determination after export. The work sits apart from ordinary export customs declaration because the goods have already left, and the question becomes who bears the cost and whether a tax rebate adjustment follows. The main data sources are customer emails, quality inspection reports, customs systems and insurer claims systems. The documents handled day to day are return agreements, claim letters, quality reports and return customs declarations. Updates arrive irregularly, case by case.
For operating reports and exception reviews, retrieval must recover the same case's subject, object, status and validity scope from these documents. A review conclusion built on a retrieved passage is only as sound as the passage's match to the case at hand. The materials jointly determine whether a return claim stands, who pays, and whether a rebate adjustment is triggered. Retrieval results feed directly into that reasoning, so the retrieval step carries the same stakes as the underlying documents.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. The vector model identifier must be checked first, and non-matching bases excluded, with the actual observation recorded.
Retrieval recalls candidates through semantic search, full-text search or hybrid search. Hybrid search merges multiple result paths through RRF. RRF is a fusion process, not a switch that can be turned on. The minimum relevance score takes effect only under semantic search or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. The reference upper limit controls the volume of cited content in tokens. Question optimization performs reference resolution and question expansion, and adds one model call before retrieval. Cited content is output as an array, and its length can be zero, so the downstream chain still runs on an empty result.
Image retrieval is not a separate capability outside the knowledge base. It adds an image input path to the existing search chain, and it depends on whether a vision model is configured or the vector model supports image input, and on whether a corresponding index was built at ingestion time. Reranking requires a model and must be enabled through settings.
Hypothetical drill sample: the source material reads, "On 10 March 2025, Company A submitted a return agreement for batch B-203; the return reason is outer packaging damage; the return declaration status is submitted but not released; the claim letter lists the claim amount and the quality report number; the insurer claims system shows pending review." The draft sentence to be judged reads, "The return declaration for this batch is complete, and the claim has been approved by the insurer."
Configuration and implementation checks
The table below lists officially documented settings or implementation check items relevant to this function, organized around the sample. The hypothetical starting values are for drill purposes only and are not official defaults.
| Configuration item | Hypothetical drill starting point | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link only one base using the same vector model | Check the vector model identifier and exclude bases on different models |
| Search mode | Hybrid search | Compare recall differences against semantic search and full-text search separately |
| Reference upper limit | Hypothetical starting point of 2000 tokens | Test content completeness with materials of different lengths, then adjust |
| Minimum relevance score | Hypothetical starting point of 0.5 | Takes effect only under semantic search or reranking; label relevant and irrelevant samples first |
| Question optimization | Not enabled for now | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Not enabled for now | Requires a model and must be enabled through settings; compare candidate ordering changes |
Three common errors
- Treating a retrieval hit as authorization or completed approval. Input is the sample source material and the draft sentence. Compare "submitted but not released" against "complete," and "pending review" against "approved by the insurer." If the draft sentence does not match the source status, the draft is judged unsupported.
- Treating the minimum relevance score as an entity or permission filter. Input is a retrieval question spanning different subjects. Compare whether returned content includes records belonging to another subject. If exclusion is attempted on similarity score alone, the understanding is wrong; retrieval does not perform authorization isolation.
- Describing RRF as a switch that can be turned on, or confusing the reference upper limit with a thinking budget. Input is the hybrid search configuration description. Compare whether an RRF switch appears, or whether the token upper limit is repurposed as a thinking budget. If either appears, the configuration description is wrong.
How to verify the setup
- Input the sample source material and the draft sentence. Compare "submitted but not released" against "complete," and "pending review" against "approved by the insurer." The condition is met if the retrieval result matches the source status and the draft sentence is marked unsupported.
- Input a linking request that spans different vector models. Compare the vector model identifiers of the bases. The condition is met if bases on different models are excluded and the actual observation is recorded. Do not run a cross-model retrieval to compare missed recalls.
- Input an explicit paper empty result []. Compare whether the downstream chain still runs. The condition is met if the empty-result chain runs and no claim is made that a fabricated rule's retrieval always returns []. These are paper checks and do not indicate actual deployment or execution.