What the documents look like
Marketplace refund and return handling generates records per order rather than per live-stream session. A refund request, a return waybill, a quality-inspection photo and a refund approval each update independently, and none of them depends on a promise made during a single broadcast. When product and marketing copy is drafted, the supporting material is scattered across the after-sales console, customer-service chat, WMS receiving and payment refund records. Conclusions about one order can sit in several separate files.
Two facts follow for knowledge-base retrieval. First, the material grows at a per-order and daily frequency. Second, the responsibility judgment, refund amount and inspection conclusion that a draft may cite often exist across documents rather than inside one. Retrieval therefore has to place order numbers, waybill numbers, inspection descriptions and approval status in a comparable position so that drafting can check them one by one. The dataset search node connects one or more knowledge bases that share the same vector model. Retrieval recalls candidates through semantic search, full-text search or hybrid search; hybrid search merges multiple result paths through RRF, and result re-ranking performs a second ordering using the ReRank score. The citation limit controls returned content length in tokens. The minimum relevance value takes effect only in semantic search or when result re-ranking is used. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires corresponding image input and model capability, and re-ranking requires a model and must be enabled in settings. The optional paths listed in the documents do not mean every request executes all of them. Retrieval itself does not perform permission filtering, a similarity score does not distinguish business entities, and it does not judge whether a record is authorized. Empty results are output as an array whose length can be 0.
Constraints at the retrieval step
A hypothetical paper exercise sample follows. The material text is: "Refund request R-2024-0517, order O-8842, buyer claims the outer packaging of the goods is damaged, requested refund amount 120 yuan, status pending review." "The return waybill shows the return for O-8842 was signed for by WMS on 20 May, inspection photo number Q-0517-03 records a compression mark on the outer carton, inner contents unopened." "Refund approval form A-0517 states an approved amount of 120 yuan, approval status pending review, no approval completion time recorded." The draft sentence to be judged is: "This order has confirmed a refund of 120 yuan, and liability for the damaged goods has been determined." What needs checking is whether the three materials contain any record of a completed refund or a liability determination.
The retrieval step does not decide those questions. It returns candidate content, and the drafting step compares that content against the draft sentence. The citation limit is measured in tokens, not in a guessed number of items, and the minimum relevance value is a filter threshold rather than a final ordering rule. After question optimization, hybrid retrieval, image retrieval or result re-ranking is enabled, the final result may pass through multi-path recall fusion and does not necessarily follow a single vector similarity ordering. A relevance score is not a category, entity or permission filter, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. Forms collect values; business verification, approval, database writes and external synchronization require their stated downstream implementation.
Configuration and implementation checks
The table below lists documented settings or implementation check items relevant to this feature and this task, around the sample above. The numeric values are hypothetical starting points for this exercise, used to compare against the actual material in the sample, and are not official default values.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | After-sales material base, copy material base | Both use the same vector model; check the vector model identifier first and exclude any that does not match |
| Search mode | Hybrid search | The sample contains an order number and natural-language description; compare semantic recall hits against full-text recall hits |
| Citation limit | Set to 1500 tokens as a hypothetical starting point | The three sample materials total several hundred characters; use it to observe whether returned content covers all three source texts, and compare completeness against a smaller value |
| Minimum relevance | Set to 0.3 as a hypothetical starting point | Used to filter low-relevance results in semantic search or result re-ranking; compare against manual annotation of relevant and irrelevant passages in the sample |
| Question optimization | Enabled | The sample question contains the reference "this order"; compare whether the subject in the original and optimized questions still points to O-8842 |
| Result re-ranking | Enabled per settings | Performs a second ordering when candidates are numerous; compare the re-ranking score against the actual relevance of the three sample materials |
Three common errors
- Treating pending review as a completed refund. In the sample, the approval status on refund approval form A-0517 is pending review, while the draft sentence states "has confirmed a refund of 120 yuan." The input is the approval form text, the comparison is the approval status field, and the observable result is that the status value does not match the draft sentence, so the sentence is unsupported.
- Treating an inspection photo as a liability determination. In the sample, inspection photo Q-0517-03 records a compression mark on the outer carton and does not record liability attribution. The input is the inspection photo description, the comparison is whether a liability determination field exists, and the observable result is that the field is absent, so the draft sentence "liability for the damaged goods has been determined" lacks a basis.
- Merging different fields into one fact. In the sample the refund request amount is 120 yuan and the approval form amount is 120 yuan, and these are different fields. The input is the two documents, the comparison is the field names and values, and the observable result is that the amounts match while the statuses differ, so a completed refund cannot be inferred from this.
How to verify the setup
- Input the three sample source texts and the draft sentence, and compare "has confirmed a refund" in the draft sentence against the approval status field on the approval form. If the returned approval form shows pending review, the draft sentence is unsupported; if the returned content does not contain the approval status, record it as an evidence gap.
- Input the sample inspection photo description and "liability has been determined" in the draft sentence, and compare whether a liability determination field exists. If the returned content contains only the outer carton compression description, the liability determination is unsupported; if the return is an empty array, handle it as an empty result and check the downstream design, without inferring that a rule necessarily returns empty.
- Input the question "what is the refund status of this order" and compare whether the subject before and after question optimization still points to O-8842. If the subject is replaced with another order after optimization, record it as a coreference error; if the optimized question matches the sample order, then compare whether the returned content covers the refund request, the return waybill and the approval form. When observing a failure branch, record the actual return, without guaranteeing a first hit or full recall.
These checks describe what to compare and what remains unknown. Without a documented error code or actual observation, the page does not assert a deterministic failure cause or result.