What the documents look like
Fresh-food weighing and sorting happens inside warehouse fulfilment, where weight and specification are the controlling facts. The working material comes from WMS, electronic scales, the order system and quality-inspection photos. The documents handled each day include sorting task sheets, weighing labels, short-weight compensation forms and quality-inspection records. These materials update per order and per day.
A single sorting task sheet can carry a Chinese product name, a foreign-language specification note and a weighing value in the same record. Cross-border and multilingual work therefore mixes natural-language description with specification codes and label codes. Semantic retrieval can match across languages, while full-text retrieval depends on keyword hits. The two paths behave differently on the same question, and that difference matters when a team needs to find a specification code or a task-sheet number rather than a description.
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 built on different vector models cannot be associated together. During implementation, check the vector model identifier first, exclude those that do not match, and record what was actually observed. Do not run a cross-model search in order to compare missed recalls.
Search mode needs to be compared across semantic, full-text and hybrid retrieval for the test question at hand. No single mode is assumed to suit all material. The citation limit is measured in tokens, so material of different lengths should be used to check whether content stays complete before the limit is adjusted. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; relevant and irrelevant samples should be labelled before the threshold is tuned. Question optimization performs coreference resolution and question expansion, so the original question and the optimized question should be compared to check whether the subject was altered.
RRF is the fusion process for mixed results and is not described as a switch. Reranking requires a model and must be enabled according to its settings. Image retrieval requires a corresponding image input and model capability. Retrieval and relevance scores do not carry permission or entity-isolation functions.
Hypothetical drill sample: the source text to be checked consists of two passages. Passage one: "Sorting task sheet SF-2024-0517, product name: chilled beef brisket, specification: 500g±10g, sorter: Li, quality-inspection status is pending inspection, weighing label printed." Passage two: "Short-weight compensation form BC-2024-0517-03, linked task sheet SF-2024-0517, requested quantity 2 units, review status is pending review, compensation reason: actual weight 460g." A draft sentence to be judged reads: "This task sheet has completed quality inspection and the compensation has been approved." Against the source text, the quality-inspection status field is "pending inspection" and the compensation form review status is "pending review." The draft sentence's claims of completed inspection and approved compensation are not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only sorting material bases that use the same vector model | Check the vector model identifier first, exclude bases on different models, record the actual association result |
| Search mode | Test semantic retrieval and full-text retrieval separately | Compare hits for specification-code questions and natural-language questions separately |
| Citation limit | Hypothetical starting point of 2000 tokens for this trial | Use sorting task sheets and compensation forms of different lengths to test content completeness, compare against a shorter limit |
| Minimum relevance | Hypothetical starting point of 0.5 for this trial | Label relevant and irrelevant samples first, compare whether retained content after filtering matches the labels |
| Question optimization | Enabled in multi-turn follow-up scenarios | Compare the original question and the optimized question, check whether task-sheet numbers and subjects were altered |
| Result reranking | Enabled according to settings when candidates are numerous | Compare ordering before and after reranking, check whether relevance scores match manual labels |
Three common errors
- Treating "pending review" as "approved." In the hypothetical sample the compensation form review status is "pending review." If the draft sentence says "compensation approved," the input is the compensation form text, the compared field is review status, and the observed result is that "pending review" and "approved" do not agree, so the claim is unsupported.
- Treating "pending inspection" as "inspection completed." In the hypothetical sample the quality-inspection status is "pending inspection." If the draft sentence says "inspection completed," the input is the task sheet text, the compared field is quality-inspection status, and the observed result is that "pending inspection" and "inspection completed" do not agree, so the claim is unsupported.
- Associating knowledge bases that use different vector models. The input is the vector model identifiers of the two knowledge bases, the comparison is whether the identifiers match, and the observed result is that the identifiers differ. The non-matching base should be excluded and the observation recorded; a cross-model search should not be run to compare missed recalls.
How to verify the setup
- For the task sheet text and the draft sentence "This task sheet has completed quality inspection and the compensation has been approved" in the hypothetical sample, the input is the two passages and the draft sentence, the comparison is against the quality-inspection status and review status fields, and the observed result is "pending inspection" and "pending review." Both claims in the draft sentence are unsupported and should be recorded as pending verification.
- For the compensation form text in the hypothetical sample, the input is the full compensation form, the comparison is against the review status field, and the observed result is "pending review." If the retrieval result returns this compensation form, check that it is still marked pending review; pending review is not treated as approval.
- For empty-result handling, the input is an explicit paper input of []. The comparison is whether the downstream chain still runs, and the observed result is that the output chain still runs when the citation content length is 0. No claim is made that a fabricated-rule search must return [].
This page covers only fresh-food weighing and sorting, cross-border material and multilingual retrieval, and knowledge base retrieval and recall. During operational handover, the authorization status of materials, operations and requests must not be treated as authorization execution through retrieval or forms.