What the documents look like
Weighing and sorting happens inside the warehouse fulfilment step, and the centre of the work is weight and specification. The main data sources are the WMS, electronic scales, the order system and quality-inspection photos. For store service and internal training, the documents that matter fall into four types: sorting task sheets, weighing labels, short-weight compensation forms and quality-inspection records. Each of these materials carries its subject, object, status and validity range on the same sheet. A sorting task sheet records the order and the specification requirement. A weighing label records the actual weight and the print time. A short-weight compensation form records the difference and the handling status. A quality-inspection record records a sampling conclusion. Updates arrive per order and per day, so retrieval has to be limited by document type and date range, otherwise sheets of the same type interfere with each other.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. Retrieval searches the whole base; a collection only groups data for management and does not change search behaviour. Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword matches. Hybrid retrieval combines the two through RRF. Re-ranking requires a re-rank model and runs when it is enabled in the settings. Image retrieval requires image input and the relevant model capability. Minimum relevance takes effect only in semantic retrieval or when result re-ranking is used. The reference limit is measured in tokens. The output reference content is an array whose length can be 0, so the downstream path still runs when nothing is found. Retrieval results do not constitute authorisation and do not handle permissions or entity isolation.
Hypothetical drill sample: a sorting task sheet reads "Task sheet no. FQ-0417-023, order tail number 8821, specification requires net weight 500 g ± 20 g, sorter A, date 17 April 2024." A weighing label reads "Label no. T-0417-023, actual weight 486 g, print time 17 April 2024, 10:12." A short-weight compensation form reads "Compensation form no. BC-0417-023, difference 14 g, status pending review, registered 17 April 2024." One draft sentence to be judged reads "This order has completed the short-weight compensation and the refund has been issued." The subject to compare is the same order tail number 8821, the object is the 14 g difference, and the status is pending review.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the sorting task sheet base that uses the same vector model | Check the vector model identifier, exclude bases with a different model, compare the association list |
| Search mode | Test semantic retrieval first, then full-text retrieval | Compare the recalled content for the same question under the two modes |
| Reference limit | Hypothetical starting point of 1500 tokens for this exercise | Compare content completeness using task sheets and compensation forms of different lengths |
| Minimum relevance | Hypothetical starting point of 0.35 for this exercise | Label relevant and irrelevant samples first, then compare the count before and after filtering |
| Question optimisation | Enable for multi-turn follow-ups, disable for a single clear question | Compare the original question with the optimised question and check whether the subject was altered |
| Result re-ranking | Enable according to settings when a re-rank model is available | Compare how re-rank scores and vector scores order the same candidate |
Three common errors
- Treating a pending short-weight compensation form as a completed refund. The input is compensation form BC-0417-023. Compare the status field "pending review" with the draft sentence "has completed the short-weight compensation and the refund has been issued." The observed result is that the status field does not state completion or refund, so the draft sentence is not supported.
- Associating knowledge bases that use different vector models and then running retrieval. The input is two bases whose vector model identifiers differ. Compare whether the identifiers match. The observed result is that a base is excluded when the identifier differs; do not run cross-model retrieval in order to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. The input is a record belonging to another store. Compare the minimum relevance setting with the source of access permission. The observed result is that a relevance value does not change the access scope, and permission has to be implemented downstream.
How to verify the setup
- Input sorting task sheet FQ-0417-023 and weighing label T-0417-023 from the hypothetical sample. Compare whether the retrieval result hits both sheets and whether the subject is the same order tail number 8821. The observed result satisfies the condition when both references appear and the subject matches.
- Input the draft sentence "This order has completed the short-weight compensation and the refund has been issued." Compare it with the status field "pending review" on compensation form BC-0417-023. The observed result is that the draft sentence is judged unsupported when the status field does not state completion or refund, and the evidence gap is retained.
- Input a paper query that clearly contains no related sheet. Compare the output reference array. The observed result is that the downstream path still runs when the array length is 0; do not claim that a fabricated rule will certainly return an empty array.