What the documents look like
Brand outlet discount operations exist to clear past-season stock, and the working document chain is made up of discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. These materials update on a weekly or monthly cycle, so a single comparison may draw on versions that span several weekly periods. The distinctive feature in this category is that price tags and approval forms may each record a discount rate, while the slow-moving list and the margin worksheet may each record a handling status for the same SKU. Subject, period and field values therefore have to be aligned item by item.
In this setting, citation tracing answers a narrow question: which material supports a given statement, what state is that material in, and does that state actually support the citation. A discount approval that is still pending review, for example, records a decision stage rather than a completed approval. A price tag print record records a printed rate rather than proof that the rate was applied at the register. These distinctions matter because price architecture damage, continued inventory buildup and margin loss are the stated consequences of getting discount operations wrong.
Constraints at the citation tracing step
The knowledge base search node can be linked to one or more knowledge bases that share the same vector model. Its output is an array of cited content, and that array may have a length of zero, meaning the output path still runs even when nothing is retrieved. Results pass through similarity filtering and a citation limit before being returned, and the minimum relevance value takes effect only under semantic retrieval or result reranking. Retrieval output is a set of citation candidates, not an authorization credential; a relevance score does not perform entity isolation or permission filtering. Hybrid retrieval merges multiple result paths through RRF, and reranking depends on model capability and on the setting being enabled. These are optional paths, and not every request executes all of them.
Hypothetical paper exercise sample: Material A is a discount approval form recording the subject as a given outlet store, the object as SKU-0417, a discount rate of sixty percent, an approval status of pending review, and a validity period from 1 March 2025 to 31 March 2025. Material B is a price tag print record for the same store and the same SKU, recording a discount rate of seventy percent and a print date of 5 March 2025. The draft sentence to be assessed reads: "SKU-0417 was executed at sixty percent in March 2025." Comparison points: the approval status is pending review, which does not constitute approval; the price tag record shows seventy percent, which is a mutually exclusive value in the same field as the sixty percent in the draft sentence; and the phrase "was executed" requires execution evidence, which neither Material A nor Material B records.
Configuration and implementation checks
The table below lists only documented settings or implementation checks relevant to the main task on this page. Numeric values are starting points for this hypothetical exercise, used for comparison against a control value, and are not official defaults.
| Configuration or check item | Hypothetical starting point | Comparison method |
|---|---|---|
| Linked knowledge bases | Select knowledge bases holding discount approval and price tag materials that share one vector model | Check the vector model identifier of each selected knowledge base, exclude those with mismatched identifiers, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare recalled content against semantic retrieval and full-text retrieval separately, and observe whether coded fields are matched |
| Minimum relevance | Set to 0.5 for this hypothetical exercise | Compare with 0.3, observe the increase or decrease in cited entries after filtering, and record the actual observation |
| Citation limit | Set to 2000 tokens for this hypothetical exercise | Compare with 1000 tokens, and observe whether the trimmed cited content covers Material A and Material B |
| Result reranking | Enabled | Compare candidate ordering against the disabled state, and confirm that a rerank model is configured |
| Citation output handoff | Connect the search node output to the answering step | Check each answer sentence against the cited source text, and confirm the downstream path still runs when the array is empty |
Three common errors
- Treating a pending review status as an approval. Material A records an approval status of pending review, while the draft sentence states that sixty percent was executed. The input is the Material A source text and the draft sentence, the comparison is between the approval status field and the execution wording, and the observable result is that the draft sentence lacks evidence of approval and of execution, so it should be recorded as unsupported.
- Treating a price tag print record as a basis for discount execution. Material B records seventy percent, while the draft sentence records sixty percent. The input is the Material B source text and the draft sentence, the comparison is between the same discount rate field, and the observable result is that the two values are mutually exclusive, which is a same-field contradiction requiring other materials for verification.
- Treating retrieval output as an authorization credential. The search node outputs a citation array whose length may be zero. The input is one retrieval output, the comparison is between the output content and authorization fields, and the observable result is that cited content only shows the material was recalled, not that the discount was approved, so authorization status still has to be checked against approval materials.
How to verify the setup
- Input the Material A source text "approval status is pending review," the Material B source text "discount rate is seventy percent," and the draft sentence "SKU-0417 was executed at sixty percent in March 2025." Compare the approval status field, the discount rate field and the execution wording. The condition is met when the comparison record shows the approval status as pending review, mutually exclusive discount rate values, and missing execution evidence, with all three retained as pending verification.
- Input a retrieval output that is an empty array. Compare whether the downstream answering step still runs. The condition is met when the downstream path is not interrupted by the empty array, and the answering step does not interpret the empty array as the material not existing or as authorization being absent.
- Input two knowledge bases whose vector model identifiers differ. Compare whether the identifiers match. The condition is met when the mismatched knowledge base is excluded, no cross-model retrieval is executed, and no inference about missed recall or redundant results is drawn from that exclusion.