What the documents look like
Refund-only dispute handling differs from ordinary return-and-refund work in one structural way: there is usually no returned parcel entering the warehouse. The evidentiary weight therefore sits on dispatch and delivery confirmation together with the value of the goods. Materials come from the platform after-sales console, carrier tracking, customer service chat, and warehouse outbound records. These materials update per order or daily.
Typical documents include the refund-only application, dispatch proof, delivery confirmation, and appeal material. The most time-consuming steps are deciding whether a return is required, collecting logistics evidence, and submitting the platform appeal. In cross-border settings the same dispute may carry a Chinese-language application alongside English-language tracking, so retrieval has to cover multilingual semantic matching.
Constraints at the retrieval step
The searchable unit is the library. A collection serves classification and management only and does not affect search results. Knowledge bases linked to a search node must share the same vector model; libraries built on different vector models cannot be linked together. The model identifier is checked first and non-matching libraries are excluded rather than searched across.
Semantic retrieval carries cross-language understanding, so a Chinese question can match an English knowledge point. Hybrid retrieval merges semantic and full-text results through RRF; RRF is a fusion process, not a standalone switch. Full-text retrieval suits order numbers, model numbers, and proper nouns. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation ceiling is measured in tokens, and token control is steadier than top k when long and short materials are mixed. Problem optimization performs coreference resolution and question expansion, which adds a model call and increases latency. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled in settings; neither runs on every request.
Hypothetical drill sample: a cross-border store's after-sales supervisor receives a refund-only application. The application text reads: "Order A-2291, item is a winter coat, customer claims non-delivery, requests refund only." The warehouse outbound record reads: "A-2291 dispatched on 2 March, carrier tracking number TEST-8842." The carrier tracking text reads: "TEST-8842 shows delivered on 6 March, signature field blank." The draft sentence awaiting judgment reads: "This order was signed for, so it does not meet refund-only conditions; recommend rejection." The comparison object is this draft sentence against the three source texts, checking whether the delivery status and the signature field support the conclusion.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Rationale and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the Chinese application library and the English tracking library that share one vector model | Check the vector model identifier first, exclude libraries on other models, and do not run cross-model retrieval |
| Search mode | Enable semantic retrieval and full-text retrieval together | Test the sample question under each mode and compare semantic hits against a full-text hit on TEST-8842 |
| Citation ceiling | Hypothetical starting value 3000 tokens | The sample holds three materials of uneven length; compare against 1500 tokens to see whether content is truncated |
| Minimum relevance | Hypothetical starting value 0.4 | Label relevant and irrelevant samples first, then compare retained volume against 0.2 and 0.6 |
| Problem optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the order number and subject were altered |
| Result reranking | Enabled when a rerank model is available | Compare candidate ordering before and after reranking and observe movement of delivery-related passages |
Three common errors
- Treating a retrieval hit as an authorized conclusion. The input is the draft sentence plus the three source texts. If the appearance of "delivered" in the tracking text alone is used to conclude that refund-only conditions are unmet, the blank signature field is being treated as a settled fact. The signature field is not recorded, so the conclusion is unsupported rather than established.
- Linking libraries built on different vector models. The input is the Chinese application library and the English tracking library. If they are merged into one search before the vector model identifier is checked, the same-model precondition stated in the documentation has been crossed, and the observed result cannot be used as grounds for claiming missed recall or redundancy.
- Treating minimum relevance as a permission filter. The input is a retrieval request containing unrelated orders. If raising the value is expected to exclude other stores or unauthorized records, similarity filtering is being confused with business authorization. Retrieval itself does not perform permission isolation.
How to verify the setup
- Input the three source texts and the draft sentence above, then compare the draft conclusion against the blank signature field in the source. If the returned citation content includes the tracking text and does not fill the blank signature field with a definite signer, the comparison condition is met. If the citation content omits the tracking text, record it as an evidence gap.
- Input the order number TEST-8842 for full-text retrieval and compare whether the returned results contain the matching outbound and tracking materials. If the order number is hit and the citation content is complete, the condition is met. If the return is an empty array, use an explicit paper input [] to check the downstream design, and do not claim that a fictional rule necessarily returns empty.
- Input a retrieval request containing only the Chinese question "has this order been signed for", and compare whether the English tracking passage returned by semantic retrieval corresponds to the question's meaning. If the returned passage contains the delivery status field and problem optimization has not rewritten the order subject, the condition is met. If the subject becomes another order after optimization, record it as pending verification and do not treat it as the cause of a retrieval failure.