What the documents look like
Fresh-food weighing and sorting happens inside warehouse fulfilment, where the working facts are weight and specification. Four document families carry the material that product and marketing copy may cite: the sorting task sheet, the weighing label, the short-weight compensation form, and the quality inspection record. These documents update per order and per day, so at drafting time the citable set is a moving collection of sheets, and any citation has to land on a specific document within a specific date range.
Material sources are the WMS, the electronic scale, the order system, and inspection photographs. A weight promise, a specification description, or a compensation statement in drafted copy can only rest on what these materials already record. Anything the materials do not record stays unknown and cannot be filled in during drafting. Errors in weighing and sorting carry consequences in short-shipment payouts, gross-margin loss, and poor customer ratings, so citation accuracy feeds directly into later fulfilment decisions.
Two features of this category shape the drafting task. Weight and specification are the core facts, which means precise identifiers such as order numbers and gram ranges matter as much as natural-language description. Documents change per order and per day, which means a citation without a date scope cannot be traced back to a version.
Constraints at the citation tracing step
Citation tracing requires that every factual statement in drafted copy correspond to source text in the materials. The knowledge base search node returns references as an array, and the length may be zero; an empty array means the search path ran through without producing references, and downstream design has to carry that case. A knowledge base search can only associate knowledge bases that share the same vector model, so cross-model knowledge bases cannot be associated together; checking starts by confirming the identifier and excluding those that do not match.
The minimum relevance value is a filter threshold. It does not perform permission or entity isolation, and semantic recall and relevance scores do not exclude unauthorized records. The same-vector-model requirement applies to knowledge base association only and does not extend to external systems. The optional paths shown in the overall retrieval diagram, such as image retrieval and reranking, are not executed on every request.
Hypothetical paper exercise. Source text one is a sorting task sheet reading: "10 March 2025, order SF-0310-017, item loose spinach, specification 500g plus or minus 20g, actual weighed 486g, sorter completed labelling." Source text two is a short-weight compensation form reading: "10 March 2025, order SF-0310-017, compensation reason short weight 14g, compensation status pending review." Source text three is a quality inspection record reading: "10 March 2025, order SF-0310-017, spinach leaves intact, no rot, inspection conclusion pass." The draft sentence to be judged is: "This order's spinach is full 500g, passed inspection, and compensation has arrived."
Configuration and implementation checks
The table below lists officially documented settings or implementation checks relevant to this feature and this task, using the sample above. Where a number appears, it is a hypothetical starting point for this exercise, with the reason for choosing it and the value it is compared against.
| Configuration or check item | Hypothetical starting point for this exercise | Reason for choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases whose vector model identifier matches the current one | Documentation states that only the same vector model can be associated together; confirm the identifier first and exclude non-matching ones |
| Reference upper limit | Set to a token amount covering the three sample source texts | Documentation controls the reference upper limit in tokens; the starting point follows the sample source length and is compared against the actual token count of the sample |
| Minimum relevance | Keep the default; do not use it to exclude unauthorized records | Documentation states the value takes effect only in semantic retrieval or with reranking, and it is a filter threshold |
| Search mode | Use full-text retrieval or hybrid retrieval for precise items such as order numbers and specifications | Documentation states full-text retrieval suits precise queries such as numbers, models, and proper nouns |
| Reference output handling | Downstream design must carry a reference array of length zero | Documentation states the reference output length may be zero; an empty array does not mean the workflow errored out |
| Document version check | Check sources against both the current and historical sorting task sheet versions | Documentation states the associated knowledge base requires a version check; whether to adopt a citation still depends on review against the valid version |
Three common errors
- Treating "compensation has arrived" in the draft sentence as a fact the materials support. In the sample, the short-weight compensation form records compensation status as pending review, and pending review does not constitute evidence of arrival. Input is the draft sentence and the compensation form text; compare the compensation status field value; the observable result is that the draft sentence wording does not match the status recorded in the material.
- Treating "full 500g" in the draft sentence as a fact the materials support. In the sample, the sorting task sheet records actual weighed 486g against a specification of 500g plus or minus 20g. Input is the draft sentence and the task sheet text; compare the actual weighed value against the specification lower bound; the observable result is that 486g falls below the lower bound of 500g minus 20g, and the draft sentence wording does not match the recorded value.
- Treating the minimum relevance setting as a means of permission or entity isolation. Input is the minimum relevance setting and the retrieval scope description; compare the documented role of that setting; the observable result is that the value is a filter threshold and does not perform the function of excluding unauthorized records, and retrieval itself does not constitute authorization.
How to verify the setup
- For the draft sentence "this order's spinach is full 500g", the input is the sorting task sheet text and the actual weighed value of 486g; compare the lower bound of the 500g plus or minus 20g specification against the measured value; the condition is met when the draft sentence wording matches the recorded value. In the current sample the two do not match, and that sentence is not supported.
- For the draft sentence "compensation has arrived", the input is the short-weight compensation form text and the compensation status field; compare the pending-review status against the arrival fact; the condition is met when the recorded status supports the arrival wording. The current sample records pending review, that sentence is not supported, and the evidence gap remains.
- For the draft sentence "passed inspection", the input is the quality inspection record text and the inspection conclusion field; compare the conclusion value under the same date, same order number, and same item; the condition is met when the recorded conclusion matches the draft sentence wording. The current sample records pass, so that sentence is supported within the sample scope, while the remaining statements still await verification.
These checks are paper judgment methods and are not claimed to have been executed. Retrieval checks record observations and failure branches only, and do not guarantee a first-hit or a complete recall. Empty-result handling can be checked on paper by supplying an empty array as input to the downstream design. Cross-model knowledge bases cannot be associated together; confirm the identifier first and exclude non-matching ones.