What the documents look like
Store count and loss control work turns on book-to-physical differences and shrinkage. It does not settle supplier payables. The working materials come from POS inventory, WMS, monitoring systems, and the SB/T counting specification. Day-to-day documents are count sheets, variance analyses, loss write-off approvals, and monitoring review records. The people involved are the loss prevention department, store inventory keepers, and financial auditors. These materials update monthly or quarterly, so a retrieval setup has to let the count sheet, the variance analysis, and the write-off approval for the same count period be recalled together for operating reports and exception review. Inventory distortion, inflated gross margin, and employee fraud are the consequences of getting this wrong. Retrieved content serves as a review lead only. It does not replace an authorization decision.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier has to be checked first, non-matching bases excluded, and the actual observation recorded. Retrieval is not authorization. Recalled content does not mean a loss write-off approval has passed or that a monitoring review has been granted.
An empty reference output is an array whose length can be zero, and the downstream chain still runs. An empty reference therefore cannot be read as a business state. The minimum relevance value is a filter threshold and takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. RRF is the fusion step for hybrid retrieval results, not a separate switch that can be turned on. Image retrieval requires image input and model capability. Reranking requires a rerank model enabled in the settings. Both are optional paths and do not run on every request.
Hypothetical paper exercise sample: a store's March 2025 count sheet reads "March 2025 count, SKU-A book quantity 120 units, physical count 112 units, variance -8 units, variance cause pending." The variance analysis for the same period reads "SKU-A variance -8 units, preliminarily judged as an unrecorded write-off, the write-off form has not been submitted for approval." The draft sentence to be judged reads "The SKU-A variance is confirmed as covered by a write-off form and no further investigation is needed." The comparison covers the subject SKU-A, the period March 2025, and the variance cause field. The source text states only that the cause is pending and that the write-off form has not been submitted for approval. The draft sentence claims confirmed coverage, which the source does not support.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Count sheet base, variance analysis base, write-off approval base | All three must share the same vector model identifier to be associated; check each identifier in turn |
| Search mode | Hybrid retrieval | Compare recall coverage against semantic retrieval alone and full-text retrieval alone |
| Reference limit | 600 tokens | Compare against 300 and 900; observe whether count sheet fields are truncated |
| Minimum relevance | 0.35 | Label relevant and irrelevant samples first, then compare against 0.2 and 0.5 |
| Question optimization | Enabled | Compare against disabled; check whether the SKU number subject is altered |
| Result reranking | Enabled | Compare against disabled; observe ranking changes when candidates are numerous |
These values are the starting point for this hypothetical trial. They are not official default values. The reference limit uses the token unit and controls the maximum amount of content cited per search.
Three common errors
- Treating a retrieval hit as a passed write-off approval. In the sample, the write-off form has not been submitted for approval. Recalling the variance analysis only shows the material exists. The approval state still needs manual verification.
- Forcing knowledge bases with different vector models into one association. The vector model identifier should be checked first, non-matching bases excluded, and cross-model retrieval should not be run to compare missed recalls.
- Treating RRF as a switch that can be turned on, or treating minimum relevance as a permission filter. RRF is the fusion step for hybrid results. Minimum relevance only filters low-relevance results and does not exclude unauthorized records.
How to verify the setup
- Input the sample source text and the draft sentence, then compare the subject, period, and field. The source supports only a pending variance cause. The draft sentence claims confirmed coverage. An unsupported claim appearing in the output meets the identification condition.
- Input the vector model identifiers for the count sheet base and the variance analysis base, then compare whether they match. When the identifiers differ, record it and exclude the non-matching base. Do not continue with a joint association.
- Input a search term containing only the SKU number, then observe the cited content and reference length. If count sheet fields are truncated, adjust the reference limit and compare again, recording the actual observation. For empty results, use the paper input [] to check the downstream design. Do not claim a fabricated rule always returns an empty array.