What the documents look like
Platform promotion registration and price management revolves around four recurring materials: the campaign registration form, the promotional price approval sheet, the inventory lock sheet, and the campaign review sheet. Compared with product title compliance work, the center of gravity here is price, stock, and campaign rhythm; promotional wording review sits in a secondary position. Data comes from platform marketing back ends, activity back ends, ERP inventory, and historical transaction prices, and it refreshes weekly or once per campaign.
During an operations handover, materials, operations, and requests each sit in a different authorization state. A retrieval result is a candidate citation. It does not constitute approval, an inventory lock, or authorization to change a price. The signature state on a promotional price approval sheet, the execution state on an inventory lock sheet, and the submission state on a registration form all have to be compared field by field against the retrieved source text. A recall score cannot stand in for that comparison.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; verify the identifiers first, exclude those that do not match, and record what was actually observed. Do not run cross-model retrieval to compare missed recalls.
Semantic retrieval depends on vector similarity. Full-text retrieval depends on keyword hits. Hybrid retrieval merges the two paths through RRF. RRF is the fusion process for hybrid results, not a switch that can be turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold, not a permission or entity-isolation mechanism. The citation limit is measured in tokens. Question optimization handles coreference resolution and question expansion and adds one model call. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and must be enabled per settings. Optional paths in the flow do not mean every request executes all of them. Citation content is output as an array and its length can be zero; the empty-result path still completes.
Hypothetical drill sample: a brand sales planning staff member intends to hand over a promotional price approval sheet. The source text reads: "Campaign name: Spring Refresh; Applicable platform: Tmall; Registered product: Model A; Promotional price: 129 yuan; Approval status is Pending review; Applicant: Sales planning role; Submission date: March 1." A second source, an inventory lock sheet, reads: "Lock target: Model A; Lock quantity: 500 units; Lock status: Locked; Lock date: March 2." The draft sentence to be judged is: "Model A's promotional price has been approved and can proceed to inventory lock and multi-platform price synchronization." When checking, compare whether the approval status field and the draft's "has been approved" refer to the same field in the same period; whether the lock status field and "can proceed to inventory lock" refer to the same object; and whether "multi-platform price synchronization" has any corresponding field in either source. It does not, so that clause is unsupported by the provided text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the promotion registration and price management base | The base must use the same vector model; verify the base identifier and exclude those that do not match |
| Search mode | Run semantic retrieval and full-text retrieval separately first | Use the same test question to compare the two recall sets, then decide whether to combine |
| Citation limit | Set to a token amount covering the full text of one approval sheet | Base it on the actual lengths of the approval sheet and inventory lock sheet, and compare whether content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic retrieval or result reranking; not used for permission filtering |
| Question optimization | Enable for handover follow-up questions that contain pronouns | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous and the question is complete | The rerank score is used for filtering; confirm the rerank model is available |
Three common errors
- Treating retrieval recall as authorized execution. Input: a pending-review approval sheet and the draft sentence. Compare the approval status field with the draft sentence. If the draft says "has been approved" while the source says "Pending review," the draft has treated a pending item as approved. Executable check: read the approval status field, compare it word by word against the draft sentence, and observe whether a state mismatch appears.
- Associating knowledge bases that use different vector models. Input: two base identifiers. Compare their vector model identifiers. If they differ and association is still attempted, the same-vector-model requirement is violated. Executable check: verify the base identifiers and model identifiers, exclude those that do not match, and record the actual observation.
- Treating an empty citation output as retrieval failure or permission denial. Input: an explicit paper empty array. Compare whether the downstream design still completes. If an empty result is judged as no permission, the output length has been confused with authorization state. Executable check: use the empty citation array to check the downstream path and observe whether it continues as designed.
How to verify the setup
- For the draft sentence about the "Pending review" approval sheet, input the approval sheet source text and the draft sentence. Compare whether the approval status field is the same field in the same period. If the source says "Pending review" and the draft says "has been approved," the draft sentence is unsupported and the evidence gap should be recorded.
- For the inventory lock sheet, input the lock status field and the draft sentence "can proceed to inventory lock." Compare whether the lock target and status refer to the same object. If the source says "Locked" while the draft says "can proceed to inventory lock," distinguish lock status from execution action and do not treat the retrieval result as execution authorization.
- For the sentence "multi-platform price synchronization," input all fields from the approval sheet and the inventory lock sheet. Compare whether any corresponding synchronization field exists. If neither source contains such a field, the sentence is unsupported, remains pending verification, and no authorization or effective clause should be supplied to fill the gap.