What the documents look like
Marketplace refund and return handling in this category occurs continuously per order, not around a single live-stream commitment. The materials arrive from the after-sales back office, customer service chat, WMS receiving, and payment refund records, and they update per order and per day. Contract and transaction material comparison at this point works across four document types: refund applications, return logistics waybills, quality inspection photos, and refund approvals. The difficult steps are deciding return responsibility, chasing the returned parcel, and approving the refund amount. Citation tracing requires that each citation map to a specific passage in a source material, and that the material version and its effective scope be checked together.
A knowledge base search node outputs a citation array whose length may be zero. An empty array means the search chain ran but matched no content; it does not mean the workflow errored out or terminated. The citation content then goes to the answering step for line-by-line comparison against the source text. Semantic retrieval recalls candidates by vector distance and forms that path's candidate order. The minimum relevance value is a filter threshold that takes effect only under semantic retrieval or when result reranking is used. Retrieval and filtering do not carry permission or entity isolation duties. Knowledge bases can be associated together only when they share the same vector model; bases with different vector models cannot be associated together, so identifiers should be checked first and non-matching bases excluded. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled according to configuration. Image retrieval requires image input and model capability. The optional paths shown in the overall flow do not all execute on every request.
Hypothetical paper exercise sample: the source material reads, "Refund application number RF-2026-0412-018, application date 12 April 2026, return logistics waybill number SF-2026-0412-7731, quality inspection photo number QC-2026-0413-009 showing outer packaging damage, refund approval status pending review." The draft sentence to be judged reads, "This order has completed refund approval and the refund has been executed." In this sample, signing, reaching agreement, and contract effectiveness are three distinct facts. Pending review is not completed approval, and it does not show that operating permission is absent.
Constraints at the citation tracing step
The search node's citation output is an array that may be empty. An empty result still lets the downstream chain run, so the receiving design must accept zero citations rather than treat that as a failure. Each citation must be traceable to a specific source passage, and the answer must be checked against that passage rather than against the fluency of the reply or the model's thinking process. The minimum relevance value filters low-relevance results only under semantic retrieval or reranking, and it is a threshold, not a final ordering rule. When question optimization, hybrid retrieval, image retrieval, or reranking is active, the final result may come from multiple recall paths merged together, so it need not follow a single vector similarity order.
Knowledge base association requires the same vector model. This requirement applies to the knowledge bases selected for one search; it does not extend to business registration sites or other external systems. Retrieval is not authorization: a returned citation does not grant permission to approve a refund, execute a payment, or synchronize an external system. Forms collect values; business verification, approval, database writes, and external synchronization require their stated downstream implementation. A pending approval status does not by itself show that the operator lacks permission. A citation array that contains content while the authorization field is empty is an evidence gap, not an approval.
Material version and effective scope must be checked alongside the citation. A current version and a historical version of the same document can both be retrieved, and whether a citation is adopted still depends on review against the effective version.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select refund material bases that match the current vector model | Compare vector model identifiers and exclude bases with a different model |
| Search mode | Hybrid retrieval | Refund numbers and natural language descriptions coexist; compare semantic and full-text recall differences |
| Citation limit | Hypothetical starting point of 3000 tokens | Estimated from average material length; compare citation completeness against 2000 and 4000 |
| Minimum relevance | Hypothetical starting point of 0.5 | Compare missed recall and irrelevant fragment ratios against 0.3 and 0.7 |
| Result reranking | Enabled | When candidates are numerous, compare ordering before and after reranking |
| Question optimization | Enabled | In multi-turn follow-up scenarios, compare retrieval results before and after completion |
These values are hypothetical starting points for this exercise, not official defaults, and need to be determined through actual comparison. The citation limit uses the documented token unit. The minimum relevance value is a 0-1 number. The search node can associate one or more knowledge bases that share the same vector model. The citation output is an array whose length may be zero.
Three common errors
- Treating a pending review status as completed approval. Input: the approval status field value "pending review." Comparison object: the value "completed." Observable result: the field values do not match, so the draft sentence is not supported by the source material.
- Treating retrieval output as authorization to execute. Input: the citation array returned by the search node. Comparison object: the business authorization status field. Observable result: citations exist while the authorization field is empty, so an evidence gap should be recorded rather than an approval.
- Associating knowledge bases with different vector models. Input: the vector model identifiers of two knowledge bases. Comparison object: whether the model names match. Observable result: the identifiers differ, so the non-matching base should be excluded and the actual observation recorded. No claim is made about what would happen if they were associated together.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare the draft phrase "has completed refund approval" against the source field "pending review." The observable result is that the two values are mutually exclusive within the same field, so the draft sentence is not supported.
- Input the citation array returned by the search node. Compare the array length against the number of source material passages. The observable result is that when the array length is zero, the downstream chain still runs, which records the empty-result receiving design. A paper input of an empty array can be used to check the downstream design; this does not claim that a fictional rule would necessarily return an empty array.
- Input the vector model identifiers of two knowledge bases. Compare whether the identifiers are the same. The observable result is that when the identifiers differ, the bases cannot be associated together; after exclusion, record the actual retrieval scope. The check stops at identifier comparison and exclusion, and does not run a cross-model retrieval to compare missed recall.
These checks state expected observations and do not claim that any test, deployment, approval, refund, or other business action has occurred. Where a document field is not provided, the value remains unknown. Where only stage-related evidence exists, the finding is unsupported or pending verification, not a contradiction.