What the documents look like
Store count and loss-control work centers on book-to-physical variance and shrinkage. It does not handle supplier settlement accounts. The working materials are count sheets, variance analyses, loss-write-off approvals, and monitoring review records. Data sources include POS inventory, WMS, monitoring systems, and count standards. Typical roles are the loss-prevention department, store inventory keeper, and finance audit.
For contract and transaction material comparison, retrieval must return records that share the same count period, the same SKU, and the same field before a quantity or status can be judged as mutually supporting. A count sheet states a book quantity and a physical quantity. A variance analysis states a cause. A write-off approval states a request number and a status. These are different fields from different stages, and a retrieval hit on one does not establish the value of another.
A hypothetical paper exercise sample is used below. It is clearly labelled and is not a real record. A supermarket store loss-prevention team checks SKU "paper goods - 3-ply facial tissue" for the March 2025 count period. Count sheet text: "Count date 2025-03-31, book quantity 120 cases, physical quantity 108 cases, variance -12 cases." Variance analysis text: "This SKU variance -12 cases, preliminarily attributed to an unrecorded write-off during the closed-store count window." Write-off approval text: "Write-off request number BX-2025-0312, quantity 12 cases, status is pending review." Draft judgement sentence: "The 12-case variance for this SKU has been fully handled by the write-off approval." The retrieval input is the above material text, and the comparison object is the write-off approval status field against the draft judgement sentence.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and is enabled according to settings. Retrieval results do not carry out authorization decisions, and a similarity score does not separate business entities or permission scopes.
The citation limit is set as tokens rather than top k because chunk lengths differ greatly across mixed knowledge bases. The minimum relevance value is a filter threshold, not a final ranking rule. When question optimization, hybrid retrieval, image retrieval, or reranking is active, the final result may come from multiple fused recall paths and does not necessarily follow a single vector similarity order. Cross-model knowledge bases cannot be associated together; the vector model identifier is checked first and non-matching bases are excluded. The optional paths shown in the overall retrieval diagram are not all executed on every request.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate the bases holding count sheets, variance analyses, and write-off approvals; confirm the vector model identifier matches | The sample materials must be jointly retrievable under one vector model |
| Search mode | Use full-text retrieval or hybrid retrieval for request number BX-2025-0312 | An exact number query depends on keyword matching |
| Citation limit | A hypothetical trial starting point of 800 tokens, chosen because the sample holds three short materials; compare against 400 tokens for content completeness | This value is a trial starting point for this exercise only |
| Minimum relevance | Annotate relevant and irrelevant samples first, then compare thresholds; effective only in semantic retrieval or reranking | Used to filter low-relevance results |
| Question optimization | Compare the original question with the optimized question; check whether subject and period were altered | The sample question contains an SKU and a date, and the subject must stay consistent |
| Result reranking | Confirm the rerank model is available and enabled according to settings | The rerank score can be used for filtering |
Three common errors
- Treating a retrieval hit as completed authorization. In the sample the write-off approval status is "pending review." Retrieving that record does not yield a conclusion that the write-off was approved or that the variance was handled. A pending request is not approval.
- Treating materials from different fields or different periods as directly contradictory. The count sheet variance of -12 cases and the write-off request quantity of 12 cases are different fields. Without mutually exclusive values in the same field, no contradiction is formed.
- Treating a similarity score as a permission or entity filter. The minimum relevance value only filters low-relevance results. It does not exclude unrelated stores or unauthorized records. Retrieval is not authorization.
How to verify the setup
- Input the draft judgement sentence and the write-off approval original text. Compare "pending review" with "fully handled." The condition is met when the judgement sentence is not supported by the original text and the pending review status remains pending review. Missing evidence remains missing.
- Input the count sheet original text and the variance analysis original text. Compare the variance quantity field with the preliminary cause field. The condition is met when the two belong to different fields and do not form a contradiction, and the variance cause remains to be verified. An unprovided field is unknown.
- Input a paper exercise with an empty result array [] and check the downstream design. Observe whether the output path still runs and whether the citation length is 0. The condition is met when the path runs through with a zero-length citation. This does not assert that a fictional rule retrieval necessarily returns []. Separately, check the vector model identifiers of the bases to be associated, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.