What the documents look like
Fresh-food weighing and sorting happens inside warehouse fulfilment. The central facts are weight and specification, not order aggregation across a pre-sale group. The working documents are the sorting task sheet, the weighing label, the short-weight compensation form and the quality inspection record. Each carries a different kind of value. A sorting task sheet identifies who weighed what, on which date, and against which specification. A weighing label records a weight value. A compensation form records a customer-reported weight gap and a status field. A quality inspection record carries a photo identifier.
For store service and internal training, these materials are used to check whether an operation followed the documented step. A training answer that cites a weight, a compensation reason or a photo number must be traceable to the specific document that carries it. The public brief describes this category as one where data updates per order and per day. That frequency does not change what citation tracing requires; it only means the cited version matters.
The eight-dimension brief is context, not a checklist. The relevant dimensions here are the document set, the actors (picker, quality supervisor, front-line staff under training), and the consequence of a wrong citation: a short-shipment payout, margin loss or a poor review. Other dimensions, such as the exact warehouse layout or the compensation approval chain, are not supplied and remain unknown.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model. Its output is a citation array, and the length may be zero. A zero-length array still lets the downstream link run. Each cited item must then be checked against the answer that uses it.
The minimum relevance setting is a filter threshold. It applies only under semantic retrieval or when result re-ranking is used. It does not act as a permission filter, an entity filter or an authorization boundary. A similarity score is not a category, entity or permission filter, and retrieval is not authorization. The same-vector-model requirement applies to linking knowledge bases; it does not extend to external systems such as a business registration site.
The overall retrieval flow lists optional paths. Image retrieval requires an actual image input and a model with the corresponding capability. Re-ranking requires a re-rank model and must be enabled by configuration. The listed paths are not all executed on every request.
A pending request is not approval. An offered option is not execution. An unsigned termination is not a completed termination. Where a document field is not provided, the value is unknown, and the citation check must record the gap rather than fill it.
Configuration and implementation checks
The table below uses a clearly labelled hypothetical paper exercise. It is not a record of a real test, deployment or approval.
Hypothetical exercise sample. Sorting task sheet A-0417 states: "17 April 2025, picker A weighed spinach, label weight 500 g, quality inspection photo number QC-0417-03." Short-weight compensation form B-0417 states: "17 April 2025, customer reported spinach actual weight 480 g, compensation requested, status is pending review." Draft sentence to be judged: "This short-weight compensation has been approved and the refund completed."
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Sorting task sheet base and quality inspection record base | Both must use the same vector model; compare model identifiers before linking |
| Search mode | Hybrid retrieval | Weight values and photo numbers need exact matching, so semantic and full-text paths run together |
| Citation limit | 800 tokens as a hypothetical starting value | Compare against a lower and a higher token setting and observe the citation array length |
| Minimum relevance | 0.4 as a hypothetical starting value | Compare with 0.3 and 0.5 and observe how many citations survive filtering |
| Question optimisation | Enabled | Compare citation content before and after enabling, for multi-turn follow-up questions |
| Result re-ranking | Enabled | Compare re-rank scores with vector scores when candidates are numerous |
These values are starting points for the paper exercise, not official defaults. RRF is a fusion step for combining multiple retrieval paths; it is not an on/off switch. Thinking budget is a model reasoning budget and must not be used as the citation token limit.
Three common errors
- Treating a retrieval result as an authorization basis. The citation array returns compensation form B-0417, whose status field reads pending review. The error is reading pending review as approved. The observable result is a draft sentence claiming approval and a completed refund when the source document shows only a pending status.
- Linking knowledge bases that use different vector models. If the sorting task sheet base and the quality inspection record base carry different vector model identifiers, they cannot be linked together. The error is skipping the identifier comparison and running retrieval across the two. The observable result is a retrieval action taken across models that the documented condition does not permit.
- Using the thinking budget as the citation token limit. The thinking budget belongs to the model reasoning configuration. The citation limit is a separate token scope on the search output. The error is mixing the two settings. The observable result is citation truncation or an unexpected citation length that does not correspond to the configured citation limit.
How to verify the setup
- Verification one. Input the hypothetical sorting task sheet A-0417 text and the draft sentence "This short-weight compensation has been approved and the refund completed." Compare the draft sentence with the compensation form B-0417 text. The form's status field reads pending review, while the draft sentence claims approval and a completed refund. The condition is met when the draft sentence conflicts with the source status field on the same subject, the same period and the same field. The result is that the draft sentence is unsupported.
- Verification two. Input the vector model identifiers of the quality inspection record base and the sorting task sheet base. Compare whether the two identifiers are the same. The condition is met when the identifiers match and linking is allowed; when they differ, exclude the non-matching base and record the gap. Do not run cross-model retrieval to compare missed recalls.
- Verification three. Input a paper exercise empty result []. Check the downstream answer stage design. Compare whether the downstream design accepts an empty array as a normal input. The condition is met when the downstream design handles an empty array without interrupting the link. This checks the documented condition that a zero-length citation array still lets the output link run. It does not assert that a real retrieval will return [].
Retrieval checks observe the failure branches. A first hit is not guaranteed, full recall is not guaranteed, and an unrelated query returning an empty array is normal. Where a document field is absent, the citation check records the gap and leaves the value unknown.