What the documents look like
Brand outlet discount operations aim to clear out-of-season inventory, and the working materials are discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. These materials are updated weekly and monthly. The same SKU can sit at different stages of the discount process across different documents, so a comparison task has to decide whether one material describes a SKU's discount in the same field, for the same period, as another material. The retrieval step only brings back relevant material fragments. Whether a discount is authorized, and whether it has taken effect in the business, still requires human checking.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together; the vector model identifier has to be checked first and non-matching bases excluded. Semantic search, full-text search and hybrid search suit different situations. Queries that turn on identifiers and model numbers are better served by full-text search or hybrid search. The reference upper limit is measured in tokens, and the minimum relevance value takes effect only under semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds an extra model call. Image retrieval needs a corresponding image input and model capability, and reranking needs a rerank model enabled under its settings. When a search returns nothing, the output path still runs, and the reference content is output as an array whose length can be 0. Retrieval and similarity scores do not perform permission or entity isolation.
Hypothetical paper exercise sample. Material A reads: "SKU-A17, March 2025 discount approval form, approval status is pending review, applicant: channel clearance group, discount rate: 60% of tag price." Material B reads: "SKU-A17, March 2025 slow-moving inventory list, handling status: listed for clearance, suggested discount: 60% of tag price, effective date: 15 March 2025." The draft sentence to be assessed is: "The discount for SKU-A17 took effect on 15 March 2025." The comparison object is the same SKU, the same period, the same discount rate field. Material A records the approval status as pending review. Material B records that the SKU is listed for clearance and gives a suggested discount and an effective date. The draft sentence presents the suggested discount and the effective date as an already effective discount, and it lacks evidence that approval was completed.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link one base containing discount approvals and slow-moving lists | Check the vector model identifier of each base and exclude bases whose model does not match |
| Search mode | Start with hybrid search | Discount approval forms carry identifiers and status words, and hybrid search covers both semantics and keywords |
| Reference upper limit | Start at a token amount that covers the complete fields of one approval form | Test with approval forms and list fragments of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Start at a middle value | Label relevant and irrelevant samples first, then compare whether key fields survive filtering |
| Question optimization | Enable it for multi-turn follow-up questions | Compare the original question with the optimized question and check whether the SKU identifier and period were altered |
| Result reranking | Enable it when candidates are numerous | Compare ordering before and after reranking and confirm the approval status field is not pushed down |
Three common errors
- Treating Material B's suggested discount and effective date as evidence that approval was completed. Material A records the approval status as pending review, and pending review does not mean approval was granted. The draft sentence's claim that the discount "took effect" lacks evidence of completed approval. Check: input the original text of Material A and Material B, compare the approval status field, and if Material A is only pending review, the draft sentence is not supported.
- Treating retrieval recall as confirmation of authorization. Retrieval only returns reference content; it does not decide whether a discount approval has been authorized. Check: input the draft sentence and the recalled fragments, compare whether a record of approval being granted appears, and if only a suggested discount is present, the authorization state still requires verification.
- Linking bases that use different vector models. Check: input the vector model identifier of each base, compare whether they match, and if they do not match, exclude that base and do not run cross-model retrieval.
How to verify the setup
- Input the complete original text of Material A and Material B together with the draft sentence. Compare the approval status field: Material A is pending review, Material B is listed for clearance. If the draft sentence claims the discount has taken effect while the approval status is still pending review, the claim is judged unsupported. Expected observation: the system recalls fragments from both materials, and human checking confirms the approval status field does not agree.
- Input a query containing only the SKU identifier, such as "SKU-A17". Compare whether the recalled results contain that SKU's approval form and slow-moving list. Expected observation: full-text search or hybrid search hits the identifier and the reference content contains the corresponding fragments; if recall is empty, check the downstream design against a paper input of an empty array, without claiming that a fabricated rule necessarily returns an empty result.
- Input a query containing an image of a discount price tag. Compare whether a vision model or a vector model that supports images has been configured, and whether an image index was built at ingestion. Expected observation: if both the model capability and the index are present, the image description or image vector participates in recall; if either condition is missing, the image retrieval path does not run, and the actual observation is recorded.