What the documents look like
Store counts and loss control work from book-to-physical variance and shrinkage. The working materials are count sheets, variance analyses, write-off approval records, and monitoring review logs. Supplier settlement sits outside this scope, so a citation path here never needs to reach accounts payable. Source data arrives from POS inventory, WMS, monitoring systems, and the SB/T counting standard, and the typical roles involved are the loss-prevention department, store warehouse keepers, and financial auditors.
Count data is refreshed monthly or quarterly, which means a citation tracing step has to keep document versions and effective periods distinct. When a training answer cites a count sheet, the citation must resolve to a specific row, a specific variance SKU, and a specific approval record. Inventory distortion, inflated gross margin, and employee fraud are the downstream consequences that make this traceability worth building. A citation that drops the material's subject, object, state, or effective range cannot support later review.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. An empty array means the search path ran through without a hit; it does not mean the workflow errored out and stopped. Cited content then needs to be passed to the answering step, where each answer sentence is checked against the cited source text.
Hypothetical drill sample: the source material reads, "March 31, 2025 closed-store count sheet, row 12, SKU 6901234567890, book inventory 48 units, physical count 43 units, variance −5 units, counter Li, reviewer Wang. This variance is listed on the April 2 variance analysis as pending investigation; the write-off approval form has not been submitted." The draft sentence under review reads, "This SKU variance has been approved for write-off and the inventory adjustment is complete." In this sample the variance is still pending investigation and the write-off approval form is unsubmitted, so the draft's approval and adjustment claims are unsupported.
Semantic recall and relevance scores affect ranking and filtering of retrieved results. They do not perform permission or entity isolation. The same-vector-model requirement is a condition for jointly associating knowledge bases; it does not extend to external systems such as business registration sites. RRF is the fusion step for mixed retrieval results, not a standalone toggle. A thinking budget is a model reasoning budget and cannot be reused as the citation token ceiling.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Count policy base, variance analysis base | Check whether the two bases report the same vector model identifier; exclude any that differ |
| Search mode | Hybrid retrieval | Compare semantic retrieval against full-text retrieval on SKU number hits |
| Citation ceiling | Hypothetical starting point 3000 tokens | Compare against count sheet paragraph length and observe whether citations are truncated |
| Minimum relevance | Hypothetical starting point 0.3 | Compare against the 0–1 rerank score range and observe whether low-score content is filtered |
| Result rerank | Enabled | Compare the ranking position of variance SKU entries before and after rerank |
| Question optimization | Disabled | Count queries have clear subjects; compare elapsed time and hits with it on and off |
The citation ceiling and minimum relevance values are hypothetical starting points for this drill, not official defaults. The configuration check must confirm whether the knowledge base search node's cited-content output is wired into the answering step.
Three common errors
- Treating a pending variance as an approved write-off. Input is the sample source text and the draft sentence; compare "the write-off approval form has not been submitted" against "has been approved for write-off." These are mutually exclusive states in the same field, so the draft contradicts the source.
- Treating an empty citation array as a workflow error. Input is the search node output
[]; compare the node's documented output definition, which permits length zero, against whether the downstream step continues. An empty array does not trigger termination. - Jointly associating knowledge bases with different vector models. Input is the vector model identifier of each knowledge base; compare whether the identifiers match. Exclude mismatches rather than running a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the sample source text and the draft sentence above. Compare the draft's "has been approved for write-off and the inventory adjustment is complete" against the source's "the write-off approval form has not been submitted." The source states no approval or adjustment fact, so the draft is judged unsupported and the evidence gap is recorded.
- Input the paper scenario where the search node outputs
[]. Compare whether the downstream answering step is designed to accept an empty array, and observe whether it continues. The expectation is that it does not terminate on an empty array. - Input the vector model identifiers of the two knowledge bases. Compare whether the identifiers match; only matching identifiers permit joint association. Exclude mismatches and record the actual observation without inferring runtime results.