What the documents look like
Weighing and sorting happen inside the warehouse fulfilment stage, and the core data is weight and specification. Four document families carry the work: sorting task sheets, weighing labels, short-weight compensation forms, and quality inspection records. They record work instructions, measurement results, difference handling, and quality determinations respectively. The material comes from the WMS, electronic scales, the order system, and inspection photographs. Updates occur per order and daily.
The task on this page is inventory procurement and supplier material. The material set covers procurement batches, supplier files, arrival acceptance records, and return grounds. Retrieval questions in this task usually carry a batch number, a supplier name, a specification grade, and a date range. Those fields are mostly numbers and proper nouns. Full-text retrieval matches numbers more directly, and semantic retrieval suits natural-language descriptions.
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 using different vector models cannot be associated together. The step is to check the vector model identifier of each knowledge base and exclude those whose identifiers do not match. This is a check on identifiers and exclusions, not an instruction to run retrieval across mismatched models and compare what comes back.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a standalone switch. Result reordering uses a ReRank model to sort candidate text a second time; it requires the model to be available and enabled according to the configuration. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. It is a filtering 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 before retrieval and adds one model call. Image retrieval requires image input and the corresponding model capability; it does not run on every request. Cited content is output as an array and its length may be 0, so the empty-result path still completes.
Hypothetical drill sample
Material text for comparison: "Supplier A, batch number A-20250410, specification grade first class, arrival date 10 April 2025, acceptance conclusion pending re-inspection. The short-weight compensation form for this batch is number B-0410-03, status pending review."
Draft sentence to be judged: "Supplier A batch A-20250410 has completed acceptance and can proceed to return settlement."
Judgement points: the material text states the acceptance conclusion is pending re-inspection, while the draft sentence states acceptance is completed. The compensation form status is pending review, while the draft sentence states return settlement can proceed. Under the same subject, the same batch, and the same field, pending re-inspection and completed acceptance are mutually exclusive, and pending review does not constitute settlement authorization. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the inventory procurement base that uses the same vector model | Check the vector model identifier of each base and exclude those that do not match |
| Search mode | Test full-text retrieval first, then hybrid retrieval | Compare the content matched by both modes using batch number A-20250410 |
| Citation limit | Start at a value that covers one complete acceptance record | Test with actual material of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic retrieval or when result reordering is used |
| Question optimization | Leave it off first and test with complete questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reordering | Enable it according to the configuration when the model is available | Compare candidate ordering before and after enabling and confirm the question is complete |
Three common errors
- Treating pending review as authorization. In the sample, compensation form B-0410-03 has status pending review. Outputting that return settlement can proceed on that basis treats pending review as completed approval. The input is the compensation form status field, the comparison is between pending review and approved, and the observable result is whether the draft sentence declares settlement executable.
- Treating pending re-inspection as completed acceptance. In the sample, the acceptance conclusion is pending re-inspection. A draft sentence stating acceptance is completed confuses mutually exclusive values in the same field. The input is the acceptance conclusion field, the comparison is between pending re-inspection and acceptance passed, and the observable result is how the draft sentence describes the acceptance status.
- Associating knowledge bases across vector models. If knowledge bases using different vector models are added to the same search node, the identifiers must be checked first and non-matching bases excluded. The input is the vector model identifier of each base, the comparison is whether the identifiers agree, and the observable result is which bases are excluded from the association list.
How to verify the setup
- Test the acceptance status with the sample material text. Input the acceptance conclusion pending re-inspection from the material text, compare it with the completed acceptance wording in the draft sentence, and if the retrieval result or generated content still shows completed acceptance, the configuration has not constrained that field.
- Test the compensation form status with the sample material text. Input the pending review status of compensation form B-0410-03, compare it with the return settlement wording in the draft sentence, and if the output does not retain the pending review limitation, check whether the relevance value and citation limit have truncated the status field.
- Check the downstream handling with an empty result. Use a clearly paper-based empty array input to check the downstream design and confirm the path still completes when the citation length is 0. This check does not claim that a hypothetical rule necessarily returns an empty array; it only observes whether the empty-result handling is usable.