What the documents look like
Brand outlet discount operations aim to clear out-of-season inventory, so the working document set is built around discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. Data comes from ERP inventory age, POS sales, outlet mall campaigns and historical discount records. In order fulfillment and after-sales question handling, customers ask about discounted prices, return and exchange conditions, and whether promotions can be combined. Each answer needs to point back to a specific document so that a price statement can be checked against an approval record or a list entry rather than restated verbally during fulfillment.
The three most time-consuming steps in this category are selecting clearance SKUs, approving discounts and checking gross margin. Citation tracing sits directly on top of those steps: the approval record carries the current status of a discount, and the slow-moving list carries the handling state of an SKU. The two are separate documents with separate fields, and neither substitutes for the other. Document updates in this category occur weekly and monthly. This page does not depend on batch synchronization or real-time triggers, so those topics are not expanded here.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of citations, and the array length may be zero, which means the output branch still runs even when no content was retrieved. The retrieved content is handed to the answering step, where each answer sentence is checked against the cited original text. The minimum relevance value takes effect only under semantic retrieval or result reranking; it is a filtering threshold and does not perform permission or entity isolation. Retrieval itself does not constitute authorization. Hybrid retrieval merges multiple retrieval paths through RRF, which is a process rather than a switch. Image retrieval requires image input and the corresponding model capability, and reranking requires a model and must be enabled by configuration; both are optional paths and do not run on every request.
Hypothetical paper exercise: On 10 March 2025, a clearance group at an outlet store submits a discount approval form. The original text reads: "Subject: clearance group, an outlet store in East China; Object: 2024 autumn/winter down jacket SKU-A, inventory age 210 days; Status: discount application pending regional manager review; Effective scope: this store only, 11 March to 20 March." A slow-moving list for the same period records the SKU as "to be handled." A draft sentence to be judged reads: "The discount for this SKU has been approved and can be sold at the new price tag immediately." The judgment must use the same subject, the same period and the same field. The approval form field is "pending review," and no approval result is recorded, so the draft sentence's "has been approved" is not supported. "Pending review" does not demonstrate a lack of operating authority, and it is not the same as approval.
Configuration and implementation checks
The table below lists documented settings or implementation check items relevant to this feature and the primary task, tied to the exercise above. The numeric values are starting points for this hypothetical exercise, used for comparison against the actual values in the sample, and are not official defaults.
| Configuration or check item | Hypothetical starting point | Comparison method and basis |
|---|---|---|
| Associated knowledge bases | Only bases containing the discount approval form and slow-moving list, sharing the same vector model | Check identifiers first; bases with different vector models are not associated together, and non-matching ones are excluded and recorded |
| Search mode | Semantic retrieval | The sample question is a natural-language follow-up, so semantic retrieval applies per the documents |
| Minimum relevance | Hypothetical starting point 0.5 | Compare against the sample retrieval score; effective only under semantic retrieval or reranking |
| Citation limit | Hypothetical starting point 2000 tokens | Compare against the actual token volume of returned citations to control citation length |
| Citation output handling | Pass the citation array to the answering step | Check each answer sentence against the cited original text |
| Citation display check | Use the approval form and list original text | Do not substitute answer fluency for source evidence |
Three common errors
- Reading "pending review" as "approved." Input is the sample approval form original text; compare the field value "pending review" with the draft sentence's "has been approved"; the observable result is that the two values are mutually exclusive, so the draft sentence is not supported.
- Treating a retrieval hit as authorization to proceed. Input is the citation array returned by retrieval; compare it with the approval status field; the observable result is that the citation only provides a source and does not change the "pending review" status.
- Treating an empty citation array as a workflow error. Input is the explicit paper output
[]; compare the downstream handling design; the observable result is that the branch still runs, and whether an error is raised depends on the downstream implementation, which the documents do not specify as an automatic error.
How to verify the setup
- Input the sample approval form original text and the draft sentence "The discount for this SKU has been approved and can be sold at the new price tag immediately." Compare the approval status field. The observable result is that "pending review" and "has been approved" are mutually exclusive, so the draft sentence is judged unsupported, and the evidence gap is recorded.
- Input the citation array returned by the sample retrieval. Compare each answer sentence with the cited original text. The observable result is that each answer can be traced back to the approval form or the list original text, with no discount conclusion lacking a source.
- Input the paper output
[]. Compare the downstream handling design. The observable result is that the branch runs and the downstream handles it as designed; do not claim that retrieval necessarily returns empty, and do not claim that an error necessarily occurs.
The documented technical facts used to identify relevant content in the documents are: the knowledge base search scope is checked first against the selected knowledge bases and document versions, and only bases with the same vector model can be associated together; the search node outputs a citation array whose length may be zero, and an empty array does not mean the workflow has errored and terminated; the retrieval output is checked for whether it is handed to the answering step, where each answer is checked against the cited original text; citation display is checked against traceable samples rather than thought processes or answer fluency; and document timeliness is checked using current and historical versions, with adoption of a citation still subject to valid-version review.