What the documents look like
Store counts and loss control work concerns book-to-physical variance and shrinkage, and it does not handle supplier payables. The working materials are count sheets, variance analyses, write-off approvals, and monitoring access records. Operating reports and exception reviews draw on POS inventory, WMS, monitoring systems, and the SB/T counting standard. Two to three facts shape this page directly.
First, the evidentiary core is a status field. A count sheet may record a variance and a linked write-off approval number, and the approval may still sit at a pending state. A variance analysis may state that the difference has been submitted for write-off approval and that approval has not been completed. These are stage statements, and they govern whether any downstream inventory action has an authorization basis.
Second, the review task is citation tracing. Every conclusion in a report or exception review must point to a specific document, field, and authorization state. A monitoring access record shows that footage was viewed on a date; it does not show that the cause of a variance was identified.
Third, the update rhythm is monthly and quarterly. Citations must carry the period they belong to, so that a March count sheet is not silently reused for a Q1 conclusion.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that share the same vector model. Its output is a citation array, and the length may be zero. A zero-length array means no content was retrieved, and the output path still runs. This matters for tracing: an empty citation set is an empty citation set, and it is not a workflow failure, an error code, or a statement that the underlying business fact is false.
The citation output must be handed to the answering step, and each answer sentence is then checked against the cited source text. The minimum relevance value is a filter threshold, and it takes effect only in semantic retrieval or when result re-ranking is used. It is not a permission filter, an entity filter, or a category filter. Retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization each need their own downstream implementation.
RRF is a fusion process for combining multiple retrieval paths. It is not a switch that is turned on. The citation upper limit is measured in tokens, not in top k. Question optimization performs coreference resolution and question expansion before retrieval, and it adds a model call.
A hypothetical paper exercise sample is set out here for checking only. A store count sheet dated March 2025 records: "Frozen goods area, shelf A3, count variance -12 units, write-off approval number BL-2025-0312, approval status is pending review." A variance analysis document states: "This variance has been submitted for write-off approval; approval has not been completed." A monitoring access record states: "12 March 2025, footage of the frozen goods area corridor was accessed; access by: loss prevention department." A draft sentence under review reads: "This variance has completed write-off approval and inventory can be written off." The draft sentence conflicts with the count sheet, which records a pending review status. The write-off approval is not complete, so no inventory write-off has an authorization basis. The monitoring record shows an access action only, and it does not show that the cause of the variance was determined; that evidence gap remains open.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Count sheet base, variance analysis base, write-off approval base | Compare the vector model identifier of the three bases; exclude any base whose identifier differs, and do not link across models |
| Search mode | Hybrid retrieval | Count sheets contain approval numbers and field names; hybrid retrieval combines semantic and keyword matching |
| Citation upper limit | Hypothetical starting point of 800 tokens | Compare against the length of a single count sheet record so that the variance SKU and approval status fields are covered |
| Minimum relevance | Hypothetical starting point of 0.3 | Compare against the semantic similarity of the write-off approval source text; content below the value is not cited |
| Question optimization | Enabled | Compare retrieval results with the setting off when a follow-up question refers to a variance SKU without naming it |
| Citation output | Array format, length zero permitted | Check whether the downstream branch still executes when the array is empty, and compare with the branch taken when citations exist |
Three common errors
- Treating a pending review status as a completed approval. Input: the count sheet field "approval status is pending review" and the draft sentence "has completed write-off approval." Compare the two status values. Pending review and completed are mutually exclusive values in the same field, so the draft sentence is not supported by the source.
- Treating a retrieval result as an authorization basis. Input: the write-off approval citation returned by the search node. Compare the citation content against the approval authorization record. The retrieval output does not contain an approval authorization, so no inventory write-off can be executed on that basis.
- Linking knowledge bases across vector models. Input: the vector model identifiers of the count sheet base and the variance analysis base. Compare whether the two identifiers are the same. When they differ, the bases cannot be linked together; exclude the non-matching base and record the observation.
How to verify the setup
- Input the count sheet text, the variance analysis text, and the draft sentence from the hypothetical sample. Compare "has completed write-off approval" in the draft sentence against "pending review" in the count sheet. The draft sentence conflicts with the source, so the judgment is that it is not supported.
- Input write-off approval number BL-2025-0312 and approval status "pending review" from the hypothetical sample. Compare the approval status against the conditions required for a write-off action. Pending review is not a completed approval, so the write-off action lacks an authorization basis.
- Input the monitoring access record "12 March 2025, footage of the frozen goods area corridor was accessed" from the hypothetical sample. Compare the access record against the time scope of the variance analysis. The access record shows an access action only, and it does not show that the cause of the variance was determined; that evidence gap stays open for further verification.