What the documents look like
Store counts and loss control examine book-to-physical differences and shrinkage. This category does not process supplier payables. The materials come from POS inventory, WMS, monitoring systems and the SB/T count standard. Day-to-day documents include count sheets, variance analyses, shrinkage write-off approvals and monitoring retrieval records. The typical roles are the loss prevention department, store warehouse keepers and financial auditors. Data updates monthly or quarterly.
When these materials enter a knowledge base, the retrieval step has to handle SKU numbers on count sheets, cause descriptions in variance analyses, status fields in write-off approvals, and time plus store identifiers in monitoring retrieval records. Numbers and statuses are exact fields; cause descriptions are natural-language fields. Both types coexist in one library, so vector model and index configuration has to serve both.
The count cycle matters for this page because a write-off approval can sit at a pending status while a count sheet already records a variance. Retrieval returns text; it does not change that status. A monitoring retrieval record documents that footage was reviewed, not that a loss event was confirmed.
Constraints at the vector indexes step
A knowledge base associated for retrieval must use the same vector model. Check the actual model and its input type before associating libraries. One source data block can map to multiple indexes, and the index content is verified separately from the source text. Build index content from complete, focused information; retrieval coverage and index length are weighed against sample material. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion, so the query side alone does not determine it. After changing a model or index, compare cited source text using the same set of questions. A vector score alone does not support a claim about outcome.
Hypothetical drill sample: a store count sheet dated 31 March 2025 reads "Store A, count date 2025-03-31, SKU number 6901234, book quantity 120 units, physical count 108 units, variance cause field empty, write-off approval status pending review." A draft sentence reads "Store A has completed write-off approval for this SKU." The comparison puts the source status field "pending review" against the draft's "completed" on the same subject, same period and same field. The two values are mutually exclusive, so the draft sentence is not supported by the source. Pending review does not show that an operator lacks permission, and it does not equal approval.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Vector model | Associate count-material libraries that share one vector model | The documents require the same vector model; verify the identifier before associating |
| Index content | Build a separate index for the variance cause field | Compare against the source block and observe whether number-type queries match |
| Index count | Keep multiple indexes for one count chunk | Compare cited source text under single-index and multi-index settings with the same questions |
| Search mode | Hybrid retrieval | Numbers use full-text, cause descriptions use semantic; compare results for both question types |
| Minimum relevance | 0.3 as the hypothetical starting point for this drill | Compare with 0.5 and observe whether low-relevance fragments are filtered |
| Citation limit | 2000 tokens as the hypothetical starting point for this drill | Compare with 1000 and observe whether cited source text is truncated |
These values are hypothetical drill starting points, not official defaults. RRF is a fusion process for mixed results, not a toggle. For empty-result handling, a paper input of [] can be used to check downstream design.
Three common errors
- Treating retrieval as authorization. The count sheet reads "write-off approval status pending review." After retrieval matches that passage, a draft sentence states "write-off approval completed." The error is treating a retrieved pending status as an executed approval. Check: input the count sheet text and the draft sentence, compare the status field values; when the source says pending review, the draft sentence does not hold.
- Associating libraries across vector models. Two count-material libraries use different vector models, and both are selected into one search node. The error is ignoring the same-vector-model requirement. Check: compare the model identifiers of both libraries; when the identifiers differ, exclude the non-conforming one and do not run cross-model retrieval to compare missed recalls.
- Treating a similarity score as a permission filter. The variance analysis contains records for Store A and Store B; retrieval returns a Store B fragment, and the operator concludes that the Store A record does not exist. The error is treating a relevance threshold as entity isolation. Check: input a Store A query, compare the store identifier in the returned fragment; when the identifier does not match, record it as a miss and do not infer that the record is absent.
How to verify the setup
- Use the hypothetical count sheet text and the draft sentence above as input to the retrieval node. Compare the status field in the returned citation against the value in the draft sentence. When the citation contains "pending review" and the draft says "completed," the draft sentence is judged unsupported and the evidence gap is recorded.
- Use SKU number 6901234 as the query with hybrid retrieval. Compare whether the returned fragment contains that number as written in the source. When the fragment contains the number and matches the count sheet, the number-type query passes; when the return is empty, record it as a miss and do not claim that a rule-based search necessarily returns [].
- Use a variance cause description as the query with semantic retrieval. Compare whether the returned fragment corresponds to the same source data block. When the fragment and source block identifiers agree, the check passes; when a different store identifier appears, record it as needing verification and do not infer that the Store A record is absent.
Image retrieval requires a corresponding image input and model capability; reranking requires a model and must be enabled in settings. The optional paths shown in the overall document flow diagram do not mean every request executes all of them. Fields not stated in the documents remain unknown, and pending review does not show that approval passed.