What the documents look like
Fresh-food weighing and sorting happens inside the warehouse fulfilment step, and the objects that matter are weight and specification. The working records come from the WMS, electronic scales, the order system and quality-inspection photos. On a daily basis the staff handle picking task sheets, weighing labels, short-weight compensation forms and inspection records, and these materials update per order and per day. Product-information and selection questions in this setting are the questions a sorter asks about specification, compensation rules or inspection decisions, and the answer needs to be traceable to the wording of a specific document. Citation tracing for this feature means every answer carries a checkable document source so the sorter can confirm whether the answer matches the currently valid version. Short shipment payouts, gross-margin loss and customer complaints are the consequences that follow when citations are inaccurate, so traceability directly affects how reliable the operation is.
Constraints at the citation tracing step
The knowledge base search node retrieves cited content from associated knowledge bases and outputs it as an array whose length may be zero. Associated knowledge bases must use the same vector model; libraries with different vector models cannot be associated together. Retrieval results pass through semantic retrieval, full-text retrieval, RRF fusion and reranking, and the combined citation content is then handed to the answering step. The minimum relevance score is a filter threshold that takes effect only under semantic retrieval or result reranking, and it does not perform permission or entity isolation. The citation limit controls how many tokens of content each search may cite. Full-text retrieval suits exact queries such as numbers, models and proper nouns, while semantic retrieval suits natural-language phrasing.
Hypothetical paper exercise: a front-warehouse inspection supervisor needs a sorter to confirm the rules for filling in a short-weight compensation form. One document in the knowledge base, marked "March 2025 version" and titled Instructions for Completing the Short-Weight Compensation Form, states: "When the actual weight recorded on the picking task sheet is lower than the order weight, the sorter fills in a short-weight compensation form, noting the order number, the actual weighed value and the difference in grams, and the inspection supervisor reviews it before submission." Another document with the same title, marked "January 2025 version", states: "When the difference exceeds 50 grams, fill in a short-weight compensation form and the sorter submits it directly." The draft sentence to be assessed is: "The short-weight compensation form is submitted directly by the sorter, with no inspection supervisor review." This draft sentence disagrees with the March version on the review step, and it also disagrees with the January version on the submitting party and the trigger condition. Which version is currently valid must be confirmed by a person before the draft sentence can be judged acceptable.
Configuration and implementation checks
The configuration is set around the checking needs of the exercise above, and only lists settings already described in the official material or implementation checks tied to this feature.
| Configuration item | Assumed starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the knowledge base containing the two compensation-form instructions | First check whether the vector model identifiers of the libraries are the same, and exclude libraries with different vector models |
| Search mode | Hybrid retrieval | The draft sentence contains proper nouns such as "short-weight compensation form" and "inspection supervisor review", so full-text retrieval can hit the keywords and semantic retrieval supplements natural-language phrasing; compare recall differences against semantic retrieval alone |
| Result reranking | Enabled | The candidate results contain two documents of close versions, and the rerank score can help distinguish relevance; compare the ordering against the run with reranking disabled |
| Citation limit | Assumed starting value for this exercise is 3000 tokens | The compensation-form instructions are short, and 3000 tokens is enough to hold the key passages of both documents; compare against a lower value to see whether the original text is truncated |
| Minimum relevance score | Assumed starting value for this exercise is 0.3 | Both version documents need to be kept for human comparison; too low a value brings in unrelated fragments and too high a value may filter out one version; compare the number of documents retained against 0.5 |
| Question optimisation | Enabled | A sorter's question may omit the subject, and question optimisation can fill in the reference; compare against the disabled run to see whether the retrieval terms include "short-weight compensation form" |
Three common errors
- Treating the citation content returned by retrieval as an authorisation basis. The knowledge base search outputs a citation array, and that array only represents the retrieved text fragments; it does not indicate that any operation has been approved. If the draft sentence is retrieved, a person still has to check the valid version of the document and the authorisation state, and the retrieval action itself does not constitute approval.
- Treating the minimum relevance score as a permission filter. The minimum relevance score is a threshold that filters out low-relevance search results, and it takes effect only under semantic retrieval or result reranking; it does not exclude records by business entity or user permission. Sorters in different roles may retrieve the same document, and permission isolation has to be implemented by a downstream system.
- Associating knowledge bases that use different vector models. Associated knowledge bases are required to use the same vector model, and libraries with different vector models cannot be associated together. If this condition is ignored and the association is made directly, the retrieval action itself is not supported, and no conclusion about missed recall or redundancy can be drawn from it.
How to verify the setup
- Enter the two compensation-form instruction texts and the draft sentence from the exercise above, and compare whether the citation content returned by retrieval contains the sentence "the inspection supervisor reviews it before submission" from the March version. If the citation content contains that sentence, the tracing path can locate a specific document; if the citation content contains only the January version or returns no compensation-form document at all, then check the vector model identifiers of the associated knowledge bases and the search mode setting.
- Enter the draft sentence "The short-weight compensation form is submitted directly by the sorter, with no inspection supervisor review", and compare whether the output of the answering step labels each citation source. If the answer cites the March version without pointing out the difference between the draft sentence and the original text on the review step, the citation display check has not passed; if the answer clearly lists the differences between the two documents and labels the versions, the tracing display meets the check condition.
- Enter a query that clearly does not concern compensation forms, such as "the print format of today's picking task sheet", and compare whether the returned citation content is an empty array. If an empty array is returned, the downstream answering step needs to handle it as designed for empty citations; if compensation-form documents are returned, then check the minimum relevance score setting and whether question optimisation introduced unrelated retrieval terms. An empty array does not mean the workflow has errored out; it only means this retrieval found no matching content.