What the documents look like
Fresh-food weighing and sorting happens inside the warehouse fulfilment step. The defining data are weight and specification, and the main sources are the WMS, electronic scales, the order system and quality-inspection photographs. The daily documents are sorting task sheets, weighing labels, short-weight compensation forms and quality-inspection records. Data update per order and per day.
At handover, each of these materials carries a different authorisation state. A sorting task sheet records the scope of work. A weighing label records the actual weighing result. A short-weight compensation form records a compensation request that is still awaiting review. A quality-inspection record records an inspection conclusion. Citation tracing has to connect each record back to the corresponding source text and version, and confirm which authorisation stage that record occupies. A record that is awaiting review is not an approved record, and an unsigned termination is not a completed termination.
Constraints at the citation tracing step
The knowledge-base search node can be linked to one or more knowledge bases that share the same vector model. Its output is a citation array, and the length may be zero. A zero-length array does not mean the workflow has errored or terminated. The retrieval step only recalls and cites content. It does not decide whether a record is authorised, and it does not isolate unrelated business or unauthorised records. The minimum relevance value is a filter threshold, not a permission filter. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. Re-ranking requires a re-rank model and must be enabled in the settings. Image retrieval requires the corresponding image input and model capability. These paths are not guaranteed to run on every request.
Hypothetical drill sample: on 10 June 2025, a front warehouse sorting task sheet reads "Task sheet no. FR-20250610-018, sorter A, item loose spinach, specification 300 g ± 20 g, planned 40 portions, work window 06:00 to 09:00." The weighing label from the same day reads "Label no. W-20250610-018-07, actual weight 286 g, printed 07:12." The short-weight compensation form from the same day reads "Compensation form no. C-20250610-003, linked label W-20250610-018-07, compensation requested 1 portion, status is awaiting quality-inspection supervisor review." The draft sentence to be assessed is "This short-weight compensation has been approved and the replacement has been completed."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the knowledge bases holding sorting task sheets, weighing labels, short-weight compensation forms and quality-inspection records | Check that the vector model identifiers of the selected knowledge bases match. Exclude those that do not match and do not link across models |
| Search mode | Hybrid retrieval | Compared with semantic retrieval and full-text retrieval, the sample contains numbered fields, and hybrid retrieval uses both semantic recall and full-text recall |
| Citation limit | Set to 3000 tokens as the starting point for this exercise | This value is the starting point for this exercise. Compare it with the total length of the sample material and adjust. It is not an official default |
| Minimum relevance | Set to 0.3 as the starting point for this exercise | This value is the starting point for this exercise. Compare it with the re-rank score and adjust. It acts only as a filter threshold |
| Result re-ranking | Enabled | Requires a re-rank model and must be enabled in the settings. Compare candidate ordering with the disabled state |
| Question optimisation | Disabled | The sample question has a clear subject. Compare elapsed time with the enabled state and decide based on actual effect |
Three common errors
- Treating retrieval output as an authorisation conclusion. The input is the short-weight compensation form text "status is awaiting quality-inspection supervisor review." The comparison object is the status field in the source material against the authorisation wording in the draft sentence. The observable result is that the two do not agree, and awaiting review does not constitute approval.
- Treating an empty citation array as an error. The input is an empty array from the search node. The comparison object is the documented statement that the length may be zero against the downstream handling design. The observable result is that the empty array is passed downstream for continued processing, and retrieval failure or business termination cannot be inferred.
- Using a similarity score in place of a permission check. The input is the minimum relevance score of a record. The comparison object is the source text and authorisation state of that record. The observable result is that the score does not change the authorisation stage of the record, and retrieval does not perform permission filtering.
How to verify the setup
- Input the sorting task sheet, weighing label and short-weight compensation form text from the hypothetical sample. Compare the draft sentence "This short-weight compensation has been approved and the replacement has been completed" against the compensation form status field "awaiting quality-inspection supervisor review." The condition is met when the draft sentence is identified as unsupported by the material, recorded as pending verification, and the evidence gap is retained.
- Input an empty array from the search node. Compare whether the downstream handling design processes a zero-length citation. The condition is met when the downstream continues execution and records an empty result, without producing an error determination.
- Input the vector model identifiers of the selected knowledge bases. Compare whether the identifiers match. The condition is met when non-matching knowledge bases are excluded and the actual observation is recorded, without running cross-model retrieval to compare missed recall.