What the documents look like
Refund-only disputes usually have no returned parcel entering the warehouse, so the evidence weight sits on dispatch, delivery signature and product value. The working materials are the refund-only application, dispatch proof, signature proof and appeal material. These materials set the definitions used in operating reports, including dispute order volume, evidence completeness rate and appeal outcome, and they determine whether a single order can be traced end to end during an exception review. Records are produced per order and per day. At the retrieval step, one order number may carry several materials at once, and status words that look similar can carry different meanings.
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 is checked first, non-matching bases are excluded, and the actual observation is recorded. Retrieval output is an array of cited content, and its length can be 0, which means the downstream path still runs when nothing is found. The minimum relevance value is a filter threshold and takes effect only under semantic retrieval or when result reranking is used; it does not perform permission or entity isolation, and retrieval itself does not constitute authorization. Hybrid retrieval merges multiple result paths through RRF. Reranking depends on a model and applies only when enabled by configuration. Image retrieval requires image input and the corresponding model capability. These paths do not run on every request.
Hypothetical paper exercise sample: the source material reads "Order A-7731, submitted by the buyer on 12 March 2025 as a refund-only application, reason: goods not received; carrier tracking shows delivery signed on 9 March, signature field blank; warehouse dispatch record states two items dispatched on 8 March." The draft sentence to be assessed is "Delivery for this order is established, so the refund-only application can be rejected directly." Assessing that sentence requires the subject, object, status and effective scope to align at the same time. A blank signature field remains unsupported, rejection is a business action, and neither retrieval nor a form produces that authority.
Configuration and implementation checks
| Configuration or check item | Value or practice on this page | Basis and comparison method |
|---|---|---|
| Linked knowledge bases | Link only dispute material bases that share one vector model | Check the vector model identifier first, exclude inconsistent bases |
| Search mode | Run the sample query once each under semantic, full-text and hybrid retrieval | Compare by test question; no single mode is presumed universally suitable |
| Citation limit | For this hypothetical trial, start at a token amount covering two materials | Chosen because the sample holds several short materials; compare completeness against longer materials |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic retrieval or result reranking |
| Question optimization | For referential questions such as "this order", compare the original and optimized question | Check whether the subject has been altered incorrectly |
| Result reranking | Compare scores only after a model is available and the setting is enabled | When not enabled, record that this item was not executed |
Three common errors
- Treating a retrieval hit as authorization. The input is the sample query, the comparison is between the cited content array and the authorization status field, and the observable result is a non-empty array with no approval or release record, so the rejection action still lacks a basis.
- Treating tracking status as an established signature. The input is the sample source text, the comparison is between the tracking status and the signature field, and the observable result is a blank signature field, so the fact remains unsupported.
- Treating pending review as approved. The input is the pending-review marker in the appeal material, the comparison is between the review status and the effective field, and the observable result is pending review only, which does not yield a completed rejection or a completed payout.
How to verify the setup
- Input the sample source text and the draft sentence, compare whether the draft sentence exceeds the subject, object, status and effective scope carried by the material, and treat a conclusion not present in the material as not satisfying the condition.
- Input an explicit paper empty result [], compare how the downstream design handles a citation length of 0, and treat continued path execution without fabricated content as satisfying the condition.
- Input two knowledge base identifiers built on different vector models, compare whether the linking configuration accepts them, and treat keeping only the same-model base while excluding and recording the rest as satisfying the condition; cross-model retrieval is not executed and its runtime result is not inferred.