What the documents look like
Central procurement and distribution for chain stores handles four document families: centralized procurement contracts, replenishment plans, distribution orders, and store requisition forms. The underlying data comes from ERP purchasing, WMS, store POS, and supplier delivery records, and it updates daily or weekly. These materials belong to three different operational stages: procurement, warehousing, and the store. The same batch of goods can carry different item codes, quantities, and date conventions across documents. When an operating report cites a source, each figure needs to trace back to a specific document and a specific field. When an anomaly review examines a variance, it needs to identify which stage produced the difference. The citation tracing step exists so that every conclusion in a report can be checked against a verifiable original text.
Constraints at the citation tracing step
Before retrieval, confirm the selected knowledge bases and the document version in use. Only knowledge bases that share the same vector model can be associated together. The dataset search node outputs cited content as an array, and the array length may be zero; an empty array does not mean the workflow has errored or terminated. Whether the cited content is passed to the answering step, and whether the answer corresponds line by line to the original text, must be checked separately. The minimum relevance score is a filtering threshold. It does not perform permission or entity isolation, and retrieval itself does not constitute authorization. A pending approval is not an approval, and an offered option is not an executed action.
Hypothetical drill sample. A chain enterprise supply chain center retains an excerpt from a centralized procurement contract: "Party A issued a replenishment plan to Party B on 1 March 2025, distribution order number PS-20250301-07, quantity 1,200 cases, delivery status signed for." A separate excerpt from a store requisition form reads: "Store M-023 submitted a requisition on 1 March 2025, item code SP-4412, quantity 300 cases, approval status pending review." The draft sentence to be judged is: "On 1 March the goods completed distribution from the distribution center to store M-023, and the store requisition was approved." Checking the two excerpts, the distribution order concerns the distribution center and the supplier, while the store requisition approval status is pending review. The draft sentence merges two different parties and two different states into one statement, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the centralized procurement contract base and the distribution order base that share one vector model | The official text requires the same vector model before association; compare against a cross-model base identifier and exclude the non-matching one |
| Citation upper limit | Start at 1,200 tokens | The official text controls the citation upper limit in tokens; compare against 800 and 1,600 and observe citation count and truncation |
| Minimum relevance score | Start at 0.4 | The official text states this takes effect only in semantic retrieval or with reranking; compare against 0.3 and 0.5 and observe the filtered count |
| Search mode | Use semantic retrieval first; switch to full-text retrieval for code-type questions | The official text states full-text retrieval suits codes and model numbers; compare recall against hybrid retrieval |
| Citation handoff | Check whether the search output array is passed into the answering node | The official text states the output is an array whose length may be zero; check the downstream design with an empty array input |
| Document version check | Retrieve against the current contract version and a historical version separately | The official text requires checking the document version; compare the two original texts, and adoption still requires review against the effective version |
Three common errors
- Treating a retrieval hit as business authorization. The search node returns a citation array, which only means content was retrieved; it does not mean a replenishment plan was approved or a distribution was completed. For the check, input the original text of the store requisition whose status is pending review, compare whether the draft sentence writes "pending review" as "approved," and observe whether the draft sentence matches the status field in the original text.
- Merging data from different parties into one conclusion. The distribution order concerns the distribution center and the supplier; the store requisition concerns the store. The objects and stages differ. For the check, extract the party, date, and quantity fields from each original text separately, compare whether the draft sentence merges facts from two parties, and observe whether the same field holds mutually exclusive values.
- Treating the minimum relevance score as a permission filter. The minimum relevance score only filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. For the check, run the same question at thresholds 0.3 and 0.5, compare the returned count and content, observe whether the threshold change alters the accessible range, and record the actual observation without inferring a permission effect.
How to verify the setup
- Citation source check: input the original text of the centralized procurement contract excerpt and the draft sentence above, compare whether the party, object, date, quantity, and status in the draft sentence match the original text, and expect the draft sentence to be marked unsupported when the status field disagrees, preserving the evidence gap.
- Empty-result handoff check: input an explicit paper empty array [], compare whether the downstream answering node still runs through, and expect the chain not to break; do not claim that a fabricated rule retrieval necessarily returns [].
- Cross-model exclusion check: first check the vector model identifiers of the two knowledge bases, compare whether the identifiers are the same, and expect that association is not performed when the identifiers differ; record only the identifiers and exclude the non-matching one, without inferring missed recall or redundant returns.