What the documents look like
Store count and loss-control work centers on book-to-physical variance and shrinkage. The materials handled are count sheets, variance analyses, write-off approvals, and monitoring access records. Supplier settlement is outside this scope. Typical data sources include POS inventory, WMS, monitoring systems, and the SB/T count standard. Loss-prevention staff, store inventory keepers, and financial auditors share these materials. Errors produce distorted inventory, inflated gross margin, or employee fraud. Data updates on monthly or quarterly cycles; this page's function does not depend on that frequency.
For contract and transaction material comparison, citation tracing requires that each citation can be returned to specific source text. The knowledge base search node selects one or more knowledge bases that share the same vector model, and outputs a citation array whose length may be zero. A zero-length array still lets the output path run through. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Minimum relevance is a filter threshold and does not perform permission or entity isolation. The same-vector-model requirement applies only to linking knowledge bases and does not extend to business-registry sites or other external systems. RRF is a multi-path result fusion process and is not an on/off switch. Thinking budget belongs to model thinking configuration and is unrelated to the citation token limit.
Constraints at the citation tracing step
A hypothetical drill sample is used here. A store count sheet dated March 2026 states: "Closing count completed; 7 SKUs show variance; write-off approval BL-0312 status is pending review; monitoring access record JK-0312 has registered a viewing window of 22:00 to 23:30." A draft sentence to be judged reads: "BL-0312 write-off has been approved, and the variance has been written off." Judgment requires the same subject, same period, and same field. The write-off approval's current status is pending review, while the draft sentence says approved. This is unsupported. Pending review does not mean the operator lacks permission, and it does not mean the variance has been written off.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the count sheet and variance analysis knowledge bases that share the same vector model | First verify the vector model identifier, exclude those with a different model, then link |
| Citation limit | 2000 tokens as the hypothetical trial starting point | Compare against variance-analysis paragraph length and observe whether citations are truncated |
| Minimum relevance | Keep default; no hypothetical adjustment in this trial | Takes effect only in semantic retrieval or result reranking; compare against full-text retrieval results |
| Search mode | Hybrid retrieval | Uses vector and full-text recall together; compare ordering after RRF fusion |
| Result reranking | Enable per settings when a rerank model is available | Compare rerank score against vector score and observe the filter effect |
| Citation output handoff | Pass the citation array to the answer step for item-by-item checking | Compare answer content against sample source text; record the gap when the array is empty |
Three common errors
- Treating pending review as approved. The input is BL-0312 with status "pending review"; the draft sentence says "approved." Compare the same field's values; they do not match, so the judgment is unsupported and the evidence gap is retained.
- Using the relevance threshold to exclude unrelated business. Minimum relevance only filters low-relevance results and does not perform permission or entity isolation. The input is a retrieval result list; compare whether an unauthorized record is present and observe whether that record still appears after the threshold is adjusted.
- Treating RRF as an on/off switch. RRF is the fusion process for hybrid retrieval results. The input is the semantic path and the full-text path; compare ordering before and after fusion and observe position changes for content hit by multiple paths.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the write-off approval's status field. The draft sentence says "approved" while the source says "pending review." The judgment is unsupported, and the evidence gap is recorded.
- Input the retrieval results and the citation array. Compare whether the answer content corresponds to source text item by item. When the citation array is empty, check the downstream handoff design. Do not claim that a fabricated rule's retrieval will necessarily return an empty array.
- Input the vector model identifiers of the linked knowledge bases. Compare whether they match, exclude those with a different model, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.