What the documents look like
Store counting and loss control work from book-to-physical differences and shrinkage. Supplier settlement is outside this scope. The materials come from POS inventory, WMS, monitoring systems, and the SB/T counting specification. Day-to-day documents include count sheets, variance analyses, damage write-off approvals, and monitoring access records. In operations permissions and handover management these materials serve one purpose: recording who holds what kind of operating eligibility over which material at which stage. A count sheet records physical and book quantities. A variance analysis records the investigation process. A write-off approval records approval status. A monitoring access record records a request for access. The retrieval step makes these materials findable. It does not decide whether a handover request has already been authorized.
Three industry facts shape this task. First, the documents carry explicit stage markers such as pending investigation and pending review, so a retrieved passage can show that a request exists without showing that anyone acted on it. Second, count sheets, variance analyses, write-off approvals, and monitoring access records are separate documents with separate fields, and the same quantity can appear in more than one of them under a different status. Third, the authorization state of a handover lives in those status fields, not in the act of retrieving the passage that mentions them.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The knowledge base is the smallest search unit; a collection is only for classifying and managing data and does not affect search results. Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval merges the two result sets through RRF. Reranking requires a ReRank model and runs according to its settings. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires a corresponding image input and model capability. The optional paths drawn in the overall retrieval diagram do not mean that every request executes all of them.
Hypothetical drill sample: a store closing count record for March 2025 (source material, two to three sentences): "After closing on 31 March 2025, storekeeper A performed a physical count of the snack area. Physical quantity 120 items, book quantity 135 items, variance -15 items, variance cause pending investigation. Loss prevention officer B received the variance analysis request on 1 April and requested access to the monitoring footage for that area from 20:00 to 22:00 on 31 March. Write-off approval form C-2025-0312 has status pending review, applicant storekeeper A, approver loss prevention supervisor, write-off quantity 15 items, write-off reason pending the variance analysis conclusion." Draft sentence to be assessed: "Storekeeper A has been granted monitoring access, the 15-item write-off has been approved, and the handover is complete." The retrieval scope is limited to the above counting and loss-control document base, and the retrieval question is "variance and write-off approval status for the snack area on 31 March."
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to this sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that use the same vector model | This sample involves only the counting and loss-control document base; verify that its vector model identifier is consistent |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | The sample contains the identifier C-2025-0312 and a date, so full-text hits on the identifier can be compared |
| Citation limit | Unit is tokens; use actual materials of different lengths to test whether content is complete | The source material contains several documents; check whether returned content covers both the variance cause field and the approval status field |
| Minimum relevance | Takes effect only in semantic retrieval or when result reranking is used | The sample contains status words such as pending investigation and pending review; label relevant and irrelevant passages before setting a threshold |
| Question optimization | Used for coreference resolution and question expansion | The sample question has a clear subject; compare whether the retrieval terms have their subject altered when the feature is on versus off |
| Result reranking | Requires a ReRank model and runs according to its settings | Candidate results may include several documents; check whether the status fields remain distinguishable after reranking |
The values above are starting points for this hypothetical trial and are not official defaults. The citation-limit starting point is chosen from the total length of the source material, and the comparison object is whether the returned content contains both the variance cause and the approval status.
Three common errors
- Treating a retrieval hit as confirmation of authorization. In the sample, the monitoring access request sits at the stage where loss prevention officer B has received the request. Retrieving that sentence shows the request exists; it does not show that storekeeper A has been granted access. The error is reading "requested access" as completed authorization.
- Reading a pending-review status as an approved write-off. Write-off approval form C-2025-0312 has status pending review, while the draft sentence states it has been approved. The error is ignoring the value of the status field and inferring an outcome from the applicant's intent.
- Mixing different fields or different periods. In the sample the variance cause is pending investigation and the write-off reason is pending the variance analysis conclusion; these are different fields. The error is treating pending investigation as confirmed shrinkage, or merging the 1 April request and the 31 March count into one status.
How to verify the setup
- Input the sample source material and the retrieval question "variance and write-off approval status for the snack area on 31 March," then compare the value of the variance cause field and the value of the approval status field in the returned content. Condition to satisfy: the returned content presents both pending investigation and pending review, and does not contain the phrase approved. If the returned content omits the approval status field, record it as not covered rather than as approved.
- Input the draft sentence "Storekeeper A has been granted monitoring access, the 15-item write-off has been approved, and the handover is complete," and compare subject, object, and status against the sample source material sentence by sentence. Condition to satisfy: identify that "has been granted access" does not match the source wording that the request was received, that "has been approved" does not match pending review, and that "handover is complete" has no corresponding record in the source. If the source does not state a handover status, record it as unsupported.
- Input the retrieval question and the limited knowledge base scope, and check whether the vector model identifiers of the associated knowledge bases are consistent. Condition to satisfy: associate only knowledge bases with the same vector model identifier, exclude those with a different identifier, and record the actual observation. If a knowledge base with a different identifier exists, do not run a cross-model retrieval to compare missed recalls; record only the exclusion. For the empty-result case, a clearly labelled paper input [] can be used to check the downstream design, without claiming that a fictional rule retrieval necessarily returns [].