What the documents look like
Brand outlet discount operations aim to clear out-of-season inventory and do not maintain a new-arrival full-price image. In store service and internal training, the cited sources center on ERP inventory age, POS sales, outlet mall campaigns and historical discounts. The day-to-day documents include discount approvals, slow-moving lists, price tags and gross-margin worksheets. These materials update weekly or monthly, so the same SKU can correspond to different approval states and price tag versions in different weeks. When a training session cites a source, the material itself, the period the material covers, and the authorization state the material currently sits in need to be recorded separately. Content retrieved by search or collected by a form only shows that the material exists; it does not show that a discount has been approved or that a price has taken effect.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model, and outputs references as an array whose length may be zero. An empty array means the search path ran through without returning content; it does not mean the workflow terminated with an error. The minimum relevance score is a filter threshold that takes effect only in semantic retrieval or when result reranking is used, and it does not serve as a permission or entity isolation function. Hybrid retrieval merges multiple result paths through RRF, reranking requires a model and is enabled according to settings, and image retrieval requires image input and model capability. The following hypothetical drill sample is used to check citations in training material.
Hypothetical drill sample: the source text reads "2025 March week 2 slow-moving list, store A, SKU-OL-331, inventory age 91 days, suggested discount 40% off, approval status is pending review." The proposed judgment sentence reads "SKU-OL-331 has received approval for a 40% discount and the price tag can be replaced." The subject and object in that sentence match the source, but the source approval status is pending review and no approval record is stated. The judgment should be treated as unsupported, and an approval document or completed approval record needs to be added. If a separate approval form for the same period and same SKU states "40% discount approved," the two materials hold mutually exclusive values on the same field, which constitutes a conflict and requires manual verification of which version is finally valid.
Configuration and implementation checks
The table below lists official documented settings or implementation check items relevant to this feature, tied to the sample above. The values are starting points for this hypothetical trial only, used for comparison with records in the sample material, and do not represent official defaults.
| Configuration or check item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only store A slow-moving list base selected | Check the vector model identifier of each selected base and exclude bases with a different vector model |
| Search mode | Hybrid retrieval | The sample contains an SKU number, so full-text and semantic recall are both needed, merged through RRF |
| Minimum relevance score | 0.5 | Hypothetical starting point, compared against each returned reference to observe whether low-relevance fragments are filtered |
| Reference limit | 800 tokens | Hypothetical starting point, compared against the actual length of the sample material to confirm whether the full text enters the reference |
| Result reranking | Enabled | The sample question is explicit, so rerank scores provide a second ordering when candidates are numerous |
| Reference output handoff | Manual verification | Check whether the reference array is passed to the answering step and compare each answer item against the source text |
Three common errors
- Treating pending review as approved. The input is the sample source text "approval status is pending review," the comparison object is "has received approval" in the proposed judgment sentence, and the observable result is that the two states do not match, so the judgment sentence is unsupported.
- Treating a retrieval hit as authorized execution. The input is the slow-moving list reference returned by search, the comparison object is the approval record required for price tag replacement, and the observable result is that the reference contains no approval field, so price tag replacement cannot be executed on that basis.
- Associating knowledge bases with different vector models together. The input is the vector model identifiers of two bases, the comparison object is the association requirement of the same vector model, and the observable result is that the identifiers differ, so that base should be excluded and recorded, with no cross-model retrieval comparison performed.
How to verify the setup
- Input the sample source text and the proposed judgment sentence, compare "pending review" against "has received approval," and observe that the result should be unsupported, with the approval record gap retained.
- Input an explicit paper empty array [], compare it against the design of the downstream answering step, and observe whether the empty reference is handled without an error, without claiming that a fictional rule search will necessarily return [].
- Input the vector model identifiers of two bases, compare whether they are identical, and observe that a base with a different identifier is excluded and recorded, without fabricating missed recall or redundant results after joint association.