What the documents look like
In refund-only dispute handling, goods frequently remain outside the warehouse, so the evidence concentrates on dispatch and delivery records together with the claimed value of the goods. The platform after-sales console, carrier tracking, customer service chat and warehouse dispatch records form the main material sources, and they update per order. When comparing contract and transaction materials, the refund-only application, dispatch proof, delivery confirmation and appeal materials need to be checked under the same subject, the same period and the same field. The retrieval step returns relevant materials as reference content. It does not perform authorization judgment or business execution.
A material set for one order typically contains short records of different types. The dispatch record states when the warehouse released the goods. The carrier tracking record states a delivery event. The refund-only application states the buyer's reason and the review status. These records may not agree with each other, and some fields may be left blank. The comparison task is to place them side by side under shared identifiers rather than to treat any single record as decisive.
Constraints at the retrieval 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. Retrieval output is an array of references, and its length can be zero. When no content is found, the output path still completes.
Semantic retrieval, full-text retrieval, hybrid retrieval and result reranking are optional paths. Reranking requires a model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. RRF is the fusion process for hybrid results and is not a standalone switch. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. It is a filtering threshold and does not constitute permission or entity isolation.
Hypothetical paper exercise sample. Material text one: "Order A-2201 was released from the warehouse on 5 March 2025; carrier tracking shows delivery on 7 March; the recipient field is blank." Material text two: "The buyer submitted a refund-only application on 9 March, stating that the goods were not received; the application status is pending review." Draft sentence to be assessed: "The order has been delivered, the buyer's claim of non-receipt contradicts the delivery record, and the refund-only application can be rejected." The fields needed for judgment are the delivery status and the application reason for the same order. A blank recipient field is an unstated item, and it cannot be used to conclude that delivery is valid or invalid.
Configuration and implementation checks
| Configuration or check item | Use on this page | Operation and observation |
|---|---|---|
| Associated knowledge bases | Limit the materials available for this comparison | Verify whether the vector model identifiers of the libraries match; exclude those that do not |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Do not assume one mode suits all materials |
| Reference limit | Unit is token; controls the amount of reference content per search | Test with materials of different lengths to see whether content remains complete |
| Minimum relevance | Takes effect only in semantic retrieval or result reranking | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Coreference resolution and question expansion | Compare the original question with the optimized question; check whether the subject was altered |
| Result reranking | Second ordering of candidate text | Confirm the model is available and enabled according to settings; observe the score range |
Hypothetical starting value for this exercise: the reference limit is set to 2000 tokens. The reason is that the sample materials contain several short passages, and a comparison with a 1000-token starting value is needed to see whether the reference content covers both the delivery and the application fields. This value is a hypothetical starting point for this exercise, not an official default.
Three common errors
- Treating a retrieval hit as an authorization conclusion. Input is the draft sentence to be assessed; the comparison object is the delivery status and application reason in the material texts; the observable result is whether the retrieved references contain both fields. If the references contain only the application reason, the delivery claim is not supported.
- Forcing knowledge bases with different vector models into one association. Input is two libraries with different vector model identifiers; the comparison object is whether the identifiers match; the observable result is whether the association operation is permitted. When identifiers do not match, exclude them and record the observation. Do not run a cross-model retrieval to compare missed recalls.
- Treating minimum relevance as a permission filter. Input is a material item unrelated to the business; the comparison object is the relevance score against the threshold; the observable result is whether that material enters the reference content. When the score is below the threshold it is filtered out, and the filtering result does not mean the material is unauthorized.
How to verify the setup
- Using material text one and material text two, enter the question "Do the delivery status and the refund-only application reason for order A-2201 agree?" Compare whether the reference content contains both the delivery record and the application reason; the observable result is that both appear and the recipient field shows blank. If only one appears, record the gap and do not judge a contradiction.
- Using the draft sentence to be assessed, enter the question "Can this order's refund-only application be rejected?" Compare whether the reference content contains authorization or approval fields; the observable result is that no such fields appear. When authorization records are missing, keep the matter pending verification and do not output a rejection conclusion.
- Using an empty reference array, check the downstream design. Compare whether the downstream handles length zero; the observable result is that the downstream path completes and generates no business action. Do not claim that retrieval under a fictional rule necessarily returns an empty array.
Unsupported, pending verification and contradictory determinations must share the same subject, the same period and the same field. A contradiction exists only when the same field has clearly mutually exclusive values. Signing, reaching cooperation and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Retrieval and form-collected values do not constitute authorized execution; approval, database writes and external synchronization require separate system implementation.