What the documents look like
Store count and loss-control material centers on book-to-physical variance and shrinkage. The main sources are POS inventory, WMS, monitoring systems and the SB/T count standard. Everyday documents include count sheets, variance analyses, damage write-off approvals and monitoring review records. These materials update monthly or quarterly, and they are largely tabular and approval-based. For one SKU, the variance conclusion, the write-off status and the review conclusion can sit in separate files.
In the order-fulfillment and after-sales answering task, retrieval has to cover variance causes, write-off approval status and monitoring review conclusions at the same time. Update frequency does not change how the search node is configured. Retrieval quality depends on chunk quality, the number of indexes and whether the search terms carry the subject.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. Cross-model retrieval is not run to compare missed recalls.
Cited content is output as an array, and its length can be 0. When nothing is found, the output path still runs. The minimum relevance value takes effect only in semantic retrieval or when result re-ranking is used. It is a filter threshold and does not perform permission or entity isolation. RRF is the fusion step for multiple result paths in hybrid retrieval; it is not a switch that can be turned on. Question optimization performs coreference resolution and question expansion and adds an extra model call.
Hypothetical paper test case
The following is a hypothetical paper test case used to check retrieval and judgment logic. It does not represent any real customer or any actual test.
Source text one (count sheet excerpt): "Store A March 2025 count sheet records SKU-0471 book quantity 120 units, physical count 108 units, variance -12 units, variance cause pending verification."
Source text two (write-off approval excerpt): "Store A March 2025 write-off form BL-0307 covers SKU-0471, requested quantity 12 units, status pending review."
Draft sentence: "The 12-unit variance for SKU-0471 at Store A has been confirmed as fully handled by write-off form BL-0307."
Comparing the draft sentence with the source texts: source one gives the variance cause as pending verification, and source two gives the status as pending review. Neither states that handling is complete. Pending review does not show a lack of operating authority, and it does not constitute a completed write-off. The judgment is unsupported, and the evidence gap is preserved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this test | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Store A count sheet base and write-off approval base, with matching vector model identifiers checked | Bases can be associated only when the vector model matches; check the identifier first, then exclude non-matching bases |
| Search mode | Hybrid retrieval for variance-cause and write-off-status queries | Compare against pure semantic and pure full-text results to see whether hits contain the subject and the status |
| Citation limit | Hypothetical starting point 3000 tokens | Test with count sheets and approval records of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Hypothetical starting point 0.35, effective only in semantic retrieval or re-ranking | Label relevant and irrelevant samples first, then compare whether status fields survive the filter |
| Question optimization | Enabled for multi-turn follow-up questions | Compare the original question with the optimized question and check whether subject and period were altered |
| Result re-ranking | Enabled when candidates are numerous | Compare ordering before and after re-ranking and confirm status fields are not pushed out by irrelevant fragments |
Three common errors
- Treating a pending-review write-off form as completed handling. The input is source text two, the compared field is status, the observed result is pending review, and the judgment should be unsupported.
- Forcing knowledge bases with different vector models into one association. The input is the vector model identifiers of two bases, the comparison is whether the identifiers match, and the observed result is that a non-matching base is excluded.
- Treating minimum relevance as a permission filter. The input is an irrelevant query and a relevant query, the comparison is the returned content after filtering, and the observed result is that the threshold affects relevance ordering only and does not exclude unauthorized records.
How to verify the setup
- Check the draft sentence against source text one. The input is the variance-cause field of the count sheet, the comparison is whether the draft sentence claims completed handling, and the observed result is that the source says pending verification, so the draft is unsupported.
- Check the write-off status against source text two. The input is the status field of the write-off form, the comparison is whether it states completion, and the observed result is pending review, so the evidence gap is preserved.
- Check the downstream design with an empty result input. The input is the explicit paper input [], the comparison is whether the output path still runs, and the observed result is that the path executes when the citation length is 0. No claim is made that a fictional rule search must return [].