What the documents look like
Fresh-food concession counters in hypermarkets are frequently run by on-site suppliers who share revenue by sales amount. Sales, shrinkage and receiving evidence are split among the store's fresh department, the concession supplier and the store manager. Cross-border material and multilingual retrieval has to work across concession contracts, daily sales reports, shrinkage sheets and food receiving inspection records. Receiving inspection records and shrinkage sheets are produced daily or per batch, and source labels and versions are easily misaligned during handover.
Citation tracing requires that an answer can return to a specific document, a specific batch and a specific clause. After retrieval recalls candidate content, the answering step has to check each citation against the source text. The citation output is an array whose length may be zero; an empty array still completes the output path, which is separate from a workflow error termination.
Constraints at the citation tracing step
Retrieval scope first checks the selected knowledge bases and the material version; only knowledge bases that share the same vector model can be associated together. A similarity score serves ranking and filtering, and does not carry permission or entity isolation. The same-vector-model requirement for associated knowledge bases constrains knowledge base association only, and does not extend to business registration sites or other external systems.
Result fusion is a combined process of multiple recall paths; RRF belongs to that process and is not a toggle. Image retrieval needs the corresponding image input and model capability, and reranking needs a ReRank model enabled by configuration. Optional paths shown in the document flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: a store keeps an excerpt of a concession contract. The source text reads: "Party A is the fresh department of a hypermarket; Party B is a concession supplier; Party B operates the leafy vegetable counter on site and shares revenue by the daily sales report amount; the contract signing date is March 1, 2025, and the effective condition is the day after both parties affix their seals." A separate food receiving inspection record reads: "April 2, 2025, the leafy vegetable counter received one batch of spinach; the supplier is Party B; the inspector is the store manager." The draft sentence to be judged reads: "The concession contract took effect on March 1, 2025, and the spinach batch was supplied directly by Party A."
Configuration and implementation checks
| Check item | Relevant setting or implementation check on this page | Explanation |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that share the same vector model | Exclude those with inconsistent identifiers first; do not run cross-model retrieval comparisons |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Exact queries such as numbers and batch codes are better served by full-text retrieval |
| Citation limit | Maximum tokens of content cited per search | This hypothetical exercise starts at 2000 tokens, chosen because it can hold several contract excerpts and inspection records; compare with 1000 tokens to observe truncation |
| Minimum relevance | A 0-1 value, effective only in semantic retrieval or with result reranking | A filtering threshold, not a means of permission or entity isolation |
| Citation content output | Array format, length may be zero | An empty array completes the output path; downstream design must accept empty citations |
| Result reranking | A ReRank model is available and enabled by configuration | When not enabled, make no assertion about rerank scores |
Three common errors
- Treating a pending review as an approval. The draft sentence states the contract took effect on March 1, 2025, while the contract source text states the effective condition is the day after both parties affix their seals, and the sample does not record a seal date. The input is the draft sentence and the contract source text; the comparison is the effective status of the same subject, same period and same field; the observable result is that the draft sentence lacks seal evidence and can only be recorded as pending verification.
- Treating a retrieval result as authorized execution. The draft sentence states the spinach batch was supplied directly by Party A, while the inspection record states the supplier is Party B. The input is the draft sentence and the inspection record; the comparison is the supplier subject field; the observable result is that the two values are mutually exclusive and the draft sentence is unsupported.
- Treating an empty citation as an error. The paper input is the citation array []; the comparison is whether the downstream node proceeds on an empty citation; the observable result is that the path continues and produces no citation content. If the downstream design aborts on an empty array, record that design gap.
How to verify the setup
- Input the contract source text and the draft sentence from the hypothetical sample, compare the effective date and effective condition fields, and observe whether the draft sentence provides a seal date. When no seal date is provided, record it as pending verification; do not conclude the contract is effective.
- Input the inspection record and the draft sentence from the hypothetical sample, compare the supplier field, and observe whether the draft sentence matches the source text. When mutually exclusive values appear, record it as unsupported.
- Input the paper citation array [], compare it against the design description of the downstream receiving node, and observe whether the path continues on an empty citation. The condition is met only when no error termination is observed; the actual runtime result remains unknown.