What the documents look like
Store counts and loss control work from book-to-physical variance and shrinkage. They do not handle supplier settlement. The main data sources are POS inventory, WMS, monitoring systems, and the SB/T count specification. Typical roles are the loss prevention department, store warehouse keepers, and financial auditors. Daily documents include count sheets, variance analyses, report-of-loss approvals, and monitoring access records.
These materials describe states, and they do not carry an authorization conclusion on their own. A count sheet records quantities and signatures. A variance analysis records a suspected cause and the status of a monitoring access request. A report-of-loss approval records a workflow stage. A monitoring access record records that a viewing action occurred. Citation tracing answers a narrower question: when a handover request cites these materials, can the material itself show which authorization state an operation or request is in. Missing material, an unstated field, or evidence that covers only an adjacent stage counts as unsupported. It does not produce a negative or conflicting conclusion.
The update rhythm is monthly and quarterly. That rhythm does not change the tracing step described here.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that share the same vector model. Its output is an array of citations, and the length may be zero. A zero-length array means the output path still runs; it does not mean the workflow errored out and stopped. The retrieval result is passed to the answering step, where each answer sentence is checked against the cited source text.
The minimum relevance value is a filter threshold. It takes effect only in semantic retrieval or when result reranking is used. It does not act as a permission filter or an entity isolation filter. Retrieval is not authorization, and a similarity score does not exclude unrelated business records or unauthorized records. Hybrid retrieval merges multiple recall paths through RRF; that merge is a process, not a switch. Image retrieval requires an image input and the relevant model capability. Reranking requires a rerank model and must be enabled by its setting. The optional paths shown in the overall flow diagram do not all execute on every request.
Hypothetical drill sample. A store's March 2025 closing count sheet states: "Aisle A, actual quantity 87 items, system book quantity 91 items, variance -4 items, counter signature Li, reviewer signature Wang." The variance analysis for the same day states: "Variance cause under investigation, monitoring access request submitted, access approval status is pending review." A draft handover sentence reads: "The Aisle A variance has received report-of-loss approval and inventory adjustment may proceed." Checking the source text: the count sheet records only quantities and signatures, and the variance analysis records only the pending investigation and the pending access review. Neither states that report-of-loss approval was granted. The draft sentence rewrites a pending-review status as granted approval, so it is unsupported.
Configuration and implementation checks
The table below stays with this sample. It lists only settings that the official documentation describes, or implementation checks tied to this page's task. The numeric values are starting points for this hypothetical trial, not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Only the store count policy base | Check the vector model identifier first, exclude bases with a different model, and record what is observed rather than running a cross-model comparison |
| Search mode | Hybrid retrieval | Item numbers and natural-language questions both appear; compare citation coverage against pure semantic retrieval |
| Minimum relevance | 0.4 | Hypothetical starting point; compare with 0.3 and 0.5 and observe whether unrelated fragments decrease |
| Citation limit | 3000 tokens | Hypothetical starting point; compare with 2000 and 4000 and observe whether citations are truncated |
| Question optimization | On | Multi-turn follow-up scenario; compare retrieval time and hits against the off setting |
| Empty-result handling | Paper input [] checked against the downstream design | Confirm the downstream design can accept a zero-length citation array; do not claim a rule returns [] |
Three common errors
- Treating pending review as granted approval. The input is the variance analysis line "access approval status is pending review." The incorrect action is to conclude that report-of-loss approval was granted. Compare the source text with the draft sentence and find that the draft adds an approval conclusion the source does not state. Observable result: the draft sentence and the source disagree on the same subject, the same period, and the same field.
- Treating a retrieval score as a permission filter. The input is the minimum relevance value of 0.4. The incorrect action is to treat records below that value as excluded because of permissions. Compare the retrieval output with the authorization record and find that the score carries no permission field. Observable result: the citation array may still contain records for unauthorized subjects, and these require separate manual review.
- Treating an empty citation array as a workflow failure. The input is a search node output of []. The incorrect action is to declare the flow errored out and stopped. Compare the node output description with the downstream configuration and find that a zero-length citation still runs the output path. Observable result: a processing gap appears only if the downstream design has no path for an empty array, and that gap must be implemented separately.
How to verify the setup
- Compare the hypothetical count sheet text "Aisle A, actual quantity 87 items, system book quantity 91 items, variance -4 items" with the draft sentence "the variance has received report-of-loss approval." The input is the two text passages; the compared field is approval status. Condition to satisfy: the draft sentence has no matching approval conclusion in the source text, so it is judged unsupported.
- Compare the hypothetical variance analysis text "access approval status is pending review" with the handover request "inventory adjustment may proceed." The input is the status field and the requested action; the compared item is the authorization boundary. Condition to satisfy: pending review does not constitute granted approval, the requested action lacks an authorization basis, and the item is judged pending verification.
- Check the downstream design against the paper input []. The input is an empty citation array; the compared item is whether the downstream node has a zero-length handling path. Condition to satisfy: if the downstream node accepts the empty array and continues, the handling design is complete; if the downstream node errors, record the gap and do not infer that a rule must return [].
The source text for these checks must sit inside an accessible target retrieval scope or be supplied as explicit manual input. When it is missing, record the gap rather than filling it.