What the documents look like
Copy drafting for fresh-food weighing and sorting draws on four record types produced inside the warehouse fulfillment step: sorting task sheets, weighing labels, short-weight compensation forms, and quality inspection records. Sorting task sheets state the order and the specification requirement. Weighing labels state the actual measured weight and the labeling result. Short-weight compensation forms state the weight difference and the handling status. Quality inspection records state the appearance judgment. These materials share one trait: they are organized around weight and specification. When draft copy uses phrases such as "full weight," "specification met," or "quality passed," the basis for those phrases sits in the records above.
The records update per order and per day. The weighing data for one batch is fixed on the same day, so the verification scope for a draft sentence can be limited to a specific date range and order range. This matters for retrieval because a draft sentence about one order should be checked against the records for that order and that day, not against the whole library.
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 node outputs reference content as an array, and the length can be 0, which means the output path still runs when nothing is retrieved. Search modes are semantic search, full-text search, and hybrid search. Hybrid search merges multiple retrieval paths through RRF. Reranking requires a model and must be enabled by configuration. The minimum relevance value takes effect only under semantic search or when result reranking is used. The reference limit is measured in tokens. Image retrieval requires the corresponding image input and model capability. The optional paths shown in the overall process diagram do not mean every request executes all of them. Retrieval is a recall action and does not perform permission or entity isolation.
Hypothetical paper exercise sample: a front warehouse sorting task sheet dated 10 June 2025 states "Order A, specification 500g±20g, sorter A." A weighing label from the same day states "Order A, actual weight 486g, labeling time 09:12." A short-weight compensation form from the same day states "Order A, difference 14g, status pending review." A draft sentence to be judged reads: "Order A has been shipped at full specification weight." Verification compares the draft sentence with the three records: the task sheet gives the specification range, the weighing label gives the actual weight, and the compensation form gives the difference status. The actual weight of 486g falls inside the 500g±20g range. The 14g difference record has a status of pending review. The phrase "shipped at full weight" in the draft sentence lacks evidence for the shipping step, and the compensation form status is pending review, which does not constitute a completed compensation or a confirmed absence of difference. That draft sentence is unsupported and needs either a shipping record or a different wording.
Configuration and implementation checks
The table below lists, for this sample, official documented settings or implementation checks relevant to this feature and this task. The values are starting points for this hypothetical exercise, used for comparison against the sample records, and are not official default values.
| Configuration item | Hypothetical exercise starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Sorting task sheet library, weighing label library | Both libraries use the same vector model. Check the library identifier before associating, and exclude any library whose vector model does not match. |
| Search mode | Hybrid search | The sample contains an exact order identifier field and a semantic specification field. Compare the semantic path and the full-text path results. |
| Reference limit | 800 tokens | The sample records are short. Use a smaller value to observe whether references stay complete, and compare against 400 and 1200 to see where content is cut. |
| Minimum relevance | 0.35 | First label relevant and irrelevant fragments in the sample, then compare the number of references after filtering at 0.2, 0.35, and 0.5. |
| Question optimization | Enabled | The sample question contains a reference. Compare the original question and the optimized question to check whether the subject was changed incorrectly. |
| Result reranking | Enabled | When candidates are numerous, compare ordering before and after reranking, and check the rerank score against the reference order. |
Three common errors
- Treating minimum relevance as a permission or entity filter. This value takes effect only under semantic search or when result reranking is used, and it is a similarity threshold. It does not exclude unrelated business records or unauthorized records. The error appears as the belief that raising this value restricts the order scope. Check method: input the sample question, compare references before and after filtering, observe whether fragments from other orders still appear, and record the actual result.
- Forcing knowledge bases with different vector models into one association. The documentation states that associated knowledge bases must use the same vector model. The error appears as associating libraries directly while ignoring the library identifier. Check method: verify the vector model identifier of each library, exclude those that do not match, and record the actual observation without inferring the runtime result of the association.
- Writing a pending-review status as a completed compensation or a confirmed absence of difference. In the sample, the short-weight compensation form status is pending review, and the draft sentence "shipped at full specification weight" lacks shipping evidence. Check method: compare the draft sentence with the status field of the compensation form, observe whether the status value and the draft wording describe the same fact, and record the unsupported part.
How to verify the setup
- Verification one: input the three sample records and the draft sentence "Order A has been shipped at full specification weight." Compare the draft sentence with the task sheet specification range, the weighing label actual weight, and the compensation form status. The condition is met when the draft wording agrees with the record fields on the same subject, the same period, and the same field. In this sample, "shipped at full weight" lacks a shipping record, and the compensation form is pending review, so the judgment is unsupported.
- Verification two: input the sample question and run semantic search, full-text search, and hybrid search separately. Compare whether the three reference sets cover the task sheet, the weighing label, and the compensation form. The condition is met when the references contain the fields needed for the judgment. If one path returns nothing, record that path as missing and do not infer that the other paths must be empty.
- Verification three: input a paper question with no corresponding record and check the output reference array. The condition is met when the output is an array of length 0, and then check whether the downstream design handles an empty result. Do not claim that a fabricated rule search will necessarily return an empty array. If the output is not empty, record the actual reference content and check its source.