What the documents look like
Warehouse picking and packing handles order outbound, and does not handle supplier arrivals. The main data sources are the WMS, the order system, the electronic waybill platform and weighing equipment. Wave picking lists, review records, waybills and packaging consumable sheets update per order or at hourly frequency. This frequency means that material relevant to copy drafting appears in new versions often, and a drafting request can reference a version that has already been replaced.
For copy drafting, the useful facts are the checkable ones: packaging consumable specifications, service marks on the waybill, and the actual operation result recorded in the review record. When drafted copy states something that does not agree with these documents, a dispute after a wrong or missing shipment, or after parcel damage, lacks traceable material support. The confidence level for this category description is broadly as stated, and the specific fields depend on the actual system output.
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 model identifier is checked first and non-matching bases are excluded. Retrieval runs against the whole knowledge base; a collection is only a classification layer and does not affect search results.
One retrieval combines question optimization, semantic retrieval, full-text retrieval or hybrid retrieval, RRF fusion and Rerank reordering, then applies similarity filtering and a reference cap. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. Reference content is output as an array, and its length can be 0, so the downstream path still runs when nothing is found. Semantic recall and relevance scores do not perform permission or entity isolation, and retrieval is not authorization. Image retrieval requires a corresponding image input and model capability, and reordering requires a model and must be enabled in the settings. Optional paths in the overall flow do not mean that every request executes all of them.
Hypothetical paper exercise sample: the source text of one review record is “12 March 2025, wave W-0312, the picker completed scanning and review of 20 items, the packaging consumable was a five-layer corrugated carton, the waybill service mark was standard express.” One draft sentence to be assessed is “This item is packed in a five-layer corrugated carton and dispatched after review on the same day.” The check is whether the subject, date, packaging material field and dispatch status are all stated in the same material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the review record base and the consumable sheet base that use the same vector model | Check the vector model identifier first, exclude bases with a different model, then observe whether recall covers the sample |
| Search mode | Hybrid retrieval | The sample contains a date and an identifier; compare whether the sample is recalled under semantic, full-text and hybrid modes |
| Reference cap | Set to a token amount that covers the full sample text | Test materials of different lengths and compare whether key fields are lost after trimming |
| Minimum relevance | Label relevant and irrelevant sample fragments first, then adjust | Takes effect only under semantic retrieval or reordering; compare whether the sample is retained before and after filtering |
| Question optimization | Enable for drafting requests that contain references | Compare the original question with the optimized question and check whether the subject was changed incorrectly |
| Result reordering | Enable according to settings when a model is available | Compare scores before and after reordering and observe whether the sample order changes |
These values are starting points for this hypothetical exercise and are not official default values.
Three common errors
- Associating knowledge bases that use different vector models. The input is two knowledge base identifiers with different vector model fields. Compare those model fields; when they differ, exclude the base. Do not run a cross-model retrieval in order to compare missed recall.
- Treating the minimum relevance value as a permission or entity filter. The input is the sample draft sentence together with an unrelated fragment. Compare the results before and after filtering; a relevance number does not exclude unrelated businesses or unauthorized records.
- Treating retrieved material as already approved copy. The input is the review record source text and the draft sentence. Compare the subject, date and status fields in both; a retrieval result does not constitute approval or authorization.
How to verify the setup
- Input the source text of the hypothetical review record and the draft sentence, run hybrid retrieval, and compare whether the recalled content contains that record. Observe whether the reference content includes the date, the packaging material and the waybill service mark. If the reference output is an empty array, record the gap; do not claim that an invented rule necessarily returns an empty result.
- Input a drafting request that contains a reference, enable question optimization, and compare the optimized result with the subject in the source text. Observe whether the subject was changed incorrectly. If the subject changes after optimization, record it as pending verification.
- Input the sample draft sentence together with a material that states only the packaging material and no dispatch status. Compare whether the dispatch wording in the draft sentence has evidence in the same field. Observe whether only a related-stage evidence exists. A field that is not stated, or evidence that covers only a related stage, is recorded as unsupported or pending verification, not as a negation or a conflict.