What the documents look like
Fresh fruit procurement for a store relies on wholesale-market price information and grade-based acceptance. The working materials are purchase orders, acceptance sheets, pesticide-residue test reports, and loss records. These are produced daily or per batch. A single batch of fruit can have its quoted price, acceptance conclusion, and test conclusion recorded by different parties at different times: a procurement clerk, a store fresh-produce supervisor, and a wholesale-market stall. The citation-tracing step takes each conclusion in an answer or handover note and points it back to the exact source text, together with the current state of that source: draft, pending review, signed, or effective. Because these materials update frequently, a version cited during handover may differ from the current version, so any citation must state the date and batch scope of the material it relies on.
A paper exercise sample is used throughout this page. It is hypothetical and does not describe any real store, supplier, or record. The sample has two source texts and one proposed handover sentence.
Source text one, a pesticide-residue test report: "Sample name: Red Fuji apple; sampling batch: 2025-06-08; test item: chlorpyrifos; test result: not detected; report status is pending review; issue date field: blank."
Source text two, a purchase order on the same page: "Supplier: a stall at Chengbei wholesale market; product: Red Fuji apple; quantity: 300 kg; unit price: to be confirmed; purchase order status: draft."
Proposed handover sentence: "The pesticide-residue test for this batch of Red Fuji apples passed, the purchase price is confirmed, and the goods can be scheduled for shelf placement."
When this sentence is checked against the two source texts, the test report status is pending review and the purchase order unit price is to be confirmed. Both conclusions in the proposed sentence lack support in the cited materials.
Constraints at the citation tracing step
Retrieval locates related content. It does not decide who may view or approve that content. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. This is an association condition. It does not extend to external business-registration systems or testing-agency systems, which are not required to use the same model.
The knowledge-base search node outputs citations as an array, and the array length may be zero. An empty array still passes through the output path; it does not mean the workflow has errored and terminated. The citation limit is measured in tokens. The minimum relevance value is a filter threshold that takes effect only under semantic retrieval or result reranking. Neither of these settings acts as a permission or entity-isolation control. Image retrieval requires image input and model capability. Reranking requires a rerank model and must be enabled by configuration. Optional paths shown in a document flow diagram do not mean every request executes all of them.
At handover, the material, the operation, and the request each carry an authorization state. Retrieval and forms are not authorization. A pending review is not an approval. An offered option is not an execution. An unsigned termination is not a completed termination.
Configuration and implementation checks
The table below ties each item to the paper exercise sample. Numeric values are hypothetical starting points for this exercise, with the reason for choosing them and the comparison value stated. They are not official defaults.
| Configuration or check item | Use on this page | Hypothetical starting value and comparison method |
|---|---|---|
| Associated knowledge bases | Limit the citation source scope | Select only knowledge bases using the same vector model; check identifiers one by one and exclude those with a different model; do not run cross-model retrieval |
| Citation limit | Control total cited tokens per search | Start at 3000 tokens, because the batch materials include a test report and a purchase order; compare citation completeness against 1500 and 6000 |
| Minimum relevance | Filter low-relevance results | Start at 0.3; compare the number of recalled passages and the proportion of unrelated fragments against 0.1 and 0.6 |
| Search mode | Determine the recall path | Use full-text or hybrid retrieval for batch numbers and identifiers; use semantic retrieval for natural-language questions |
| Citation output | Carry an empty result | Use a paper input of [] to check whether the downstream node continues; do not claim that a particular rule necessarily returns [] |
| Manual check action | Check material state | Compare each answer conclusion against the source text on subject, object, state, and effective scope |
Three common errors
- Treating a retrieval hit as a permission or effectiveness judgment. Input the test report whose status is pending review. Compare whether the answer writes "test passed" while the source says "pending review." Observe whether the answer preserves the pending-review state. A similarity score does not establish authorization or effectiveness.
- Treating an empty citation array as a workflow failure. Input a paper empty array
[]. Compare whether the downstream node continues to execute. Observe whether an unconfigured error branch appears. An empty array passing through the output path is existing node behavior. - Treating a quoted price validity period as a supply period. Input the purchase order draft whose unit price field says "to be confirmed." Compare whether the handover sentence writes "price confirmed." Observe whether the conclusion exceeds what the material states. An unstated field remains unsupported.
How to verify the setup
- Input the hypothetical test report text and the proposed handover sentence. Compare the two field values "pending review" and "test passed." Observe whether the answer rewrites pending review as passed. The condition is met when the answer preserves pending review and marks the evidence gap.
- Input the hypothetical purchase order draft and the phrase "price confirmed." Compare the unit price field "to be confirmed" with "confirmed." Observe whether the answer declares the price confirmed. The condition is met when the answer states that the unit price lacks support in the material.
- Input the paper empty array
[]and a downstream answer node. Compare whether the node continues to execute. Observe whether the output path passes through. The condition is met when the downstream continues and no unconfigured error assertion is produced.
These checks describe expected observations. They do not report that any test, deployment, approval, or business action has taken place. Where the source material does not state a field, the gap remains open rather than being resolved by assumption.