What the documents look like
Trade returns and claims handling covers reverse logistics and responsibility determination after export. Source material arrives from customer emails, quality inspection reports, customs systems and insurer claim systems. The working documents for a case are the return agreement, the claim letter, the quality report and the return customs declaration. In store service and internal training, these materials must be retrieved, cited and turned into a handover version that staff can check line by line.
Document updates are irregular and case by case. Each return and claim matter produces its own set of files, and a retrieval configuration does not depend on a fixed synchronisation rhythm. The practical consequence for this page is that retrieval quality rests on the completeness and labelling of each case file, not on a schedule.
The four document types differ in what they can answer. A return agreement records the parties, the goods and the reason for return. A claim letter records what the claimant asserts and what has been submitted. A quality report records inspection findings. A return customs declaration records the declaration status. Retrieval can surface any of these, and the retrieved citation still has to be compared against the source wording before it becomes a handover statement.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Search mode is compared across semantic retrieval, full-text retrieval and hybrid retrieval according to the test question. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process, not a separate switch that can be turned on. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimisation performs coreference resolution and question expansion and adds an extra model call. Retrieval returns citation content as an array whose length can be zero, which means the downstream path still runs when nothing is found. Retrieval itself does not perform authorisation checks, and the minimum relevance value is a filtering threshold rather than a permission or entity isolation mechanism.
A clearly labelled hypothetical paper exercise sample follows. The source text for checking is: "On 10 March 2025, return agreement number TR-0310 was signed by Party A and Party B, covering 12 industrial pump bodies, with bearing noise as the reason for return; the return customs declaration has been submitted, and the claim letter has not yet been submitted." The draft sentence to be assessed is: "This return has completed claim submission and can proceed to the compensation process." The comparison checks the draft sentence against the fields in the source text. The source states only that the return customs declaration has been submitted and that the claim letter has not yet been submitted. The draft sentence records claim submission as completed, which is not supported by the source text.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases that use the same vector model | The return agreement, claim letter, quality report and return customs declaration must be retrieved under one vector model |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Number TR-0310 and "bearing noise" exercise the full-text and semantic paths respectively |
| Citation limit | Measured in tokens; test completeness with real materials of differing length | The return agreement and the quality report differ in length, so observe whether citations are truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic retrieval or result reranking, filtering low-relevance fragments |
| Question optimisation | Compare the original question with the optimised question and check whether the subject was altered | On a follow-up question about the status of this return, check whether the number and subject are preserved |
| Result reranking | Used when a rerank model is available and enabled by configuration | Secondary ordering of text results when candidates are numerous, producing a score from 0 to 1 |
The citation limit value used in this exercise is a hypothetical starting point chosen because the return agreement and quality report in the sample differ in length; it is compared against the observed completeness of retrieved citations, and it is not an official default. The minimum relevance value is likewise a hypothetical starting point, selected after labelling relevant and irrelevant fragments in the sample, and compared against the labelled set.
Three common errors
- Treating the draft sentence "this return has completed claim submission" as a fact from the source text. The source states only that the claim letter has not yet been submitted. The comparison must proceed field by field, and a status that is not recorded is marked as unsupported rather than written as a denial or a conflict.
- Treating the minimum relevance value as a permission or entity isolation setting. That value is a filtering threshold and takes effect only under semantic retrieval or result reranking. Retrieval does not determine whether a store staff member is authorised to view a particular case.
- Associating knowledge bases that use different vector models. Before association, check the vector model identifier and exclude those that do not match. Do not run retrieval across differing vector models in order to compare missed recalls.
How to verify the setup
- Enter the hypothetical sample source text and the draft sentence, compare the field "the claim letter has not yet been submitted" with the field "claim submission completed", and observe whether the two are identified as mutually exclusive values of the same field. The condition is met when the draft sentence is judged unsupported and the evidence gap around the unsubmitted claim letter is preserved.
- Enter the number TR-0310 and run full-text retrieval and semantic retrieval separately, compare whether the returned citations contain that number and the reason-for-return field, and observe whether the citation content can be checked against the source. The condition is met when the citation content contains the source subject, object and status. A first-position hit or complete recall is not required.
- Enter the paper input [] for an empty result, check whether the downstream node still runs, and observe whether a path with a citation length of zero appears. The condition is met when the downstream design can accept empty citations. Do not claim that a fictional rule returns [] in every case, and do not claim that an unrelated query must return nothing.