What the documents look like
Fresh-food concession counters in hypermarkets are frequently run by on-site supplier staff under a sales-share arrangement. Stock procurement and supplier records therefore span the store fresh-food department, the concession supplier and the store manager. The documents handled on this page are the concession contract, the daily sales report, the shrinkage record and the food purchase inspection record. Data sources include the POS system, supplier delivery notes, food-safety ledgers and store stock counts, with updates occurring daily or per batch. These materials serve three purposes at once: food-safety inspection, gross-profit sharing and shrinkage allocation. Errors can trigger food-safety penalties, gross-profit disputes and customer claims.
Citation tracing addresses a specific need: keeping the original source of each document retrievable and traceable through the search and answering chain. The daily update frequency and per-batch scope matter because a citation must point to the version that was actually in effect at the time of the transaction, not to a later revision. The three-party structure matters because a citation may need to attribute a statement to the store, the supplier or the manager, and the source document determines which party's record supports it.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its output is an array of citations whose length may be zero; even when no content is retrieved, this output path still completes. The retrieval chain includes semantic search, full-text search, hybrid search, result reranking, question optimization and image retrieval. These paths are triggered in combinations determined by the input content and search parameters, and not every request executes all of them.
The citation limit is measured in tokens. The minimum relevance score is a 0–1 filter threshold that takes effect only in semantic search or when result reranking is used. This threshold is a filter value; it does not perform permission or entity isolation, and retrieval itself does not constitute authorization.
Hypothetical paper exercise: a store fresh-food department intends to check one concession-counter food purchase inspection record. The source text has three sentences: "Supplier A delivered one batch of chilled pork to this store's fresh-food concession counter on 10 March 2025, accompanied by delivery note number SD-0310-07. The store inspected and recorded the purchase inspection information for this batch on the same day, record number JY-20250310-02. The shrinkage record for this batch has not yet been submitted, and the shrinkage allocation amount is pending confirmation." The draft sentence to be assessed is: "The purchase inspection record for this batch has been fully archived, and the shrinkage allocation has been completed." Compared with the source text, the draft sentence matches the subject and object, but the shrinkage allocation status in the source text is pending confirmation, and the claim that it has been completed is not supported.
Configuration and implementation checks
The table below lists documented settings or implementation checks relevant to this function and primary task, using the sample above. The value column gives a hypothetical starting point for this exercise only, with the reason for choosing it and the comparison object.
| Configuration item | Setting or check action in this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Include only knowledge bases confirmed to use the same vector model | Knowledge bases with different vector models cannot be associated together; verify identifiers first and exclude those that do not match |
| Search mode | Use hybrid search, with RRF fusing semantic and full-text results | Identifier-type content such as SD-0310-07 depends on full-text recall, while natural-language questions depend on semantic recall |
| Minimum relevance score | Hypothetical starting point of 0.3, compared with the actual score of the sample source text | This value takes effect only in semantic search or result reranking and is a filter threshold |
| Citation limit | Hypothetical starting point of 2000 tokens, compared with the length of the sample source text | This value controls the total amount of cited content per search and is not counted in items |
| Citation output handoff | Check whether the search node output is passed to the answering step | Compare each answer against the cited source text to confirm the source can be traced back |
| Citation display check | Use the sample source text and draft sentence to check the displayed content | Do not substitute answer fluency for source evidence |
Three common errors
- Treating the minimum relevance score as a permission or entity isolation mechanism. This value is only a 0–1 filter threshold that takes effect in semantic search or result reranking. It does not exclude unrelated businesses or unauthorized records, and retrieval does not constitute authorization.
- Converting the token unit of the citation limit into an item count. The documents control the citation limit in tokens and provide no conversion rule to item counts. Setting it by item count is an unsupported inference.
- Writing a pending status as completed. In the sample, the shrinkage allocation amount is pending confirmation, while the draft sentence states it has been completed. This is unsupported. Where a field is not stated or only related-stage evidence exists, it should be recorded as pending verification.
How to verify the setup
- Input the sample source text and the draft sentence, then compare the draft claim "the shrinkage allocation has been completed" with the source statement "the shrinkage allocation amount is pending confirmation." If the claim is judged unsupported, this meets the expectation. This is a paper check and has not been run.
- Input the sample delivery note number SD-0310-07 and compare the recall behavior of full-text search and semantic search, observing whether identifier-type content is matched through the full-text path. If only the semantic path matches or nothing matches, record the actual observation without presupposing a failure cause.
- Use a paper input of an empty citation array to check the downstream handoff design, confirming that the output path still completes with an empty array and that the empty array is not judged as a workflow error termination. This check is at the design level and does not claim that a fabricated rule search necessarily returns an empty array.
Knowledge bases with different vector models cannot be associated together. During checks, only verify identifiers and exclude those that do not match; do not run cross-model retrieval to compare missed recalls. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled per the settings. The optional paths shown in the overall document flow diagram do not mean that every request executes all of them.