What the documents look like
Refund-only dispute work draws on four material families: the platform after-sales backend, carrier tracking, customer service chat, and warehouse outbound records. The distinguishing feature of this category is that a refund-only dispute frequently has no returned parcel entering the warehouse, so the evidence emphasis falls on dispatch proof and signed-delivery proof rather than on inbound inspection. Copy drafting that describes item condition, delivery facts, or value therefore has to be anchored to those two documents in particular.
Material updates at per-order and daily frequency mean that a draft citation for one order has to bind to the currently effective version of that order's evidence. A sentence drafted today against a dispatch record may no longer match the record after the order's evidence set changes.
The drafting step compares each proposed sentence against these materials one by one, confirming whether a promotional or factual claim has a corresponding source. The knowledge base search module can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. Search mode, question optimization, result reranking, and the citation limit together shape the final citation set; hybrid search merges multiple retrieval paths through RRF. The minimum relevance value is a filtering threshold that takes effect only in semantic search or when result reranking is used, and it does not perform permission or entity isolation. The retrieval step itself completes no authorization decision and executes no business action.
Hypothetical paper exercise. An after-sales specialist handles a refund-only request for order HYP-2025-0310-A on 10 March 2025, concerning a portable speaker labelled with water resistance grade IPX4. Three source sentences are available for comparison. The dispatch record states: "This order left the East China warehouse on 3 March 2025, item model HYP-SPK-01." The carrier tracking states: "Signed for on 6 March 2025, signatory shown as the recipient in person." The customer service chat states: "Customer reports the speaker will not power on after water entered it, and requests a refund only." One draft sentence is pending judgement: "This model has been laboratory tested and works normally in rain; after signing, the customer caused water ingress through personal misuse, so responsibility does not rest with the merchant." The phrases "laboratory tested" and "responsibility does not rest with the merchant" have no corresponding source in the materials, and "works normally in rain" does not form an explicit mutually exclusive value against the IPX4 label in the same field, so the sentence is unsupported. The handover request is: please review whether this draft sentence can serve as an item description in appeal materials.
Constraints at the citation tracing step
The search module's citation output is an array, and a zero-length array still lets the downstream chain run. An empty array is not evidence that the workflow has errored or terminated. Citation tracing for copy drafting requires a human comparison of each answer sentence against the cited source text, checking whether the subject, object, status, and effective scope stated in the draft appear in the materials.
Hybrid search merges semantic and full-text recall through RRF; RRF is a fusion process for combined results, not a switch to be turned on. Reranking requires a rerank model and must be enabled under its settings; image retrieval requires corresponding image input and model capability. The optional paths shown in the overall retrieval flow are not all executed on every request.
The minimum relevance value filters low-relevance results and does not act as a permission, entity, or business filter. Retrieval is not authorization. A relevance score is not a category, entity, or permission filter, and no similarity setting excludes unrelated businesses or unauthorized records.
Configuration and implementation checks
The table below is tied to the hypothetical exercise above. Numeric entries are starting points for this paper exercise only, each with its reason and comparison object.
| Configuration or check item | Starting point in this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the already-chosen after-sales knowledge base | Compare the vector model identifier of each library; only libraries with the same vector model may be associated together. Exclude and record libraries with a different identifier. |
| Search mode | Hybrid search | The sample contains precise terms such as model HYP-SPK-01 and IPX4; compare recalled content against a semantic-search-only result. |
| Citation limit | Set to 2000 tokens as the starting point for this exercise | This is a trial starting value; compare against the token volume of the three source sentences plus the draft sentence, then adjust. |
| Minimum relevance | Set to 0.5 as the starting point for this exercise | This is a trial starting value; it takes effect only in semantic search or with result reranking. Compare against the rerank score. |
| Question optimization | Not enabled in this exercise | The sample question already contains the order number and model; enabling it adds one model call. Compare against elapsed time before deciding. |
| Citation handover | Pass search output to the copy answer step | Compare each answer sentence against the cited source text; treat a zero-length array as an empty result for downstream design review. |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Input: a retrieval request against the same knowledge base in the sample. Comparison object: citation content belonging to different orders. If an operator assumes that raising the threshold excludes unrelated orders, that is a misreading. The minimum relevance value only filters low-relevance results and does not isolate business entities.
- Treating a pending draft sentence as already authorized. Input: the sentence "responsibility does not rest with the merchant" in the sample. Comparison object: the customer service chat and the dispatch record, neither of which contains a responsibility-determination field. A pending-review status does not show that the sentence has been approved, and does not show that the merchant lacks operating authority.
- Forcing association of knowledge bases with different vector models. Input: two libraries with different vector model identifiers. Comparison object: the association condition. If the identifier difference is ignored and the libraries are associated directly, that is an incorrect operation. The documentation states only that libraries with the same vector model may be associated together; it does not describe the runtime behaviour after such an association.
How to verify the setup
- Input the three source sentences and the draft sentence from the sample. Compare the phrases "laboratory tested" and "responsibility does not rest with the merchant" against the material fields. If the materials contain no corresponding source, judge the sentence unsupported and record the evidence gap.
- Input a retrieval request for order HYP-2025-0310-A. Compare the returned citation array against the three source sentences. If the array contains the dispatch record and the signed-delivery proof text, the source-traceability condition is satisfied. If the array length is zero, handle it as an empty result when checking the downstream design, and do not claim that retrieval necessarily returns empty.
- Input two knowledge bases with different vector model identifiers. Compare the identifier fields. If the identifiers differ, exclude the joint association, record only the observed identifier difference, and do not assert missed recall or redundant results from cross-model retrieval.