What the documents look like
Concession counters in shopping malls operate under the mall's unified cashiering and deduction settlement. Brand ERP, mall POS, membership campaign records and the finance system each hold their own entries. In cross-border work, the joint-operation deduction contract, mall campaign letter, sales receipt and monthly reconciliation statement frequently appear in different languages, and a search must map a question asked in one language onto source text written in another. Daily and monthly materials coexist, and citation tracing requires that an answer can return to the specific source text so that the subject, period and field can be checked.
Two or three industry facts shape this task directly. Because the mall handles cashiering centrally, the sales figures a brand sees in its own ERP may not match the mall POS record, so a citation must point to the exact document and field being reconciled. Because these documents arrive in multiple languages, the retrieval step has to support cross-language matching rather than keyword identity. Because campaign letters and deduction contracts carry different states, a citation must preserve whether a document is an invitation, a signed confirmation or a completed settlement.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. Citation content then passes to the answering step, where each answer statement is checked against the cited source text. Semantic retrieval supports cross-language understanding. The minimum relevance value takes effect only under semantic retrieval or result reranking, and it is a filtering threshold. Retrieval itself does not perform authorization judgment.
Hypothetical paper exercise sample. The source text to be checked is: "Joint-operation deduction contract (hypothetical number H-01), clause 3: the counter's monthly sales for Party B are confirmed against the mall's unified cashiering records, and the deduction ratio is twenty percent of the month's tax-inclusive sales. The contract term runs from the hypothetical date 1 January 2025 to 31 December 2025." A second document reads: "Mall campaign letter (hypothetical number H-02): the counter is invited to join the hypothetical March 2025 promotion; costs are to be discussed separately; the reply slip deadline is 20 February 2025." The draft sentence to be judged is: "The counter has confirmed participation in the March promotion and agreed to bear the campaign costs." The check compares the invitation wording against the draft. The campaign letter only issues an invitation, the reply deadline has not passed, and no signature or confirmation record exists. The phrases "has confirmed participation" and "agreed to bear the costs" have no basis in the materials and are unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the contract base and campaign letter base whose vector model identifier matches | Compared with a base whose vector model identifier differs; mismatched bases are excluded and no cross-model retrieval is performed |
| Search mode | Semantic retrieval, for cross-language questions | Compared with full-text retrieval; exact queries for numbers and models use full-text or hybrid retrieval |
| Minimum relevance | Hypothetical starting value 0.3 | Compared with 0.5 to observe change in recall count; this value takes effect only under semantic retrieval or reranking |
| Citation limit | Hypothetical starting value 2000 tokens | Compared with 1000 to observe whether citation content is trimmed |
| Question optimization | Enabled for multi-turn follow-up questions | Compared with disabled, to observe coreference resolution and added latency |
| Result reranking | ReRank enabled when candidates are numerous | Compared with not enabling it, to observe the 0-1 score ordering |
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is the campaign letter text and the draft sentence; the comparison is between "invitation" and "has confirmed participation"; the observable result is that the draft sentence lacks signature or reply-slip evidence and is marked unsupported.
- Treating minimum relevance as a permission or entity filter. The input is a question spanning several counters; the comparison is whether the returned citations include other counters' records; the observable result is that relevance affects only ordering and filtering and does not isolate subjects, so access scope must be checked separately.
- Treating an empty citation array as a workflow error. The input is the paper value []; the comparison is whether the downstream node continues to execute; the observable result is that a citation length of zero is permitted and an empty array does not equal workflow termination.
How to verify the setup
- Input the contract text above together with the draft sentence. Compare whether the draft sentence contains confirmation or cost-bearing wording not present in the source. The condition is met when the draft sentence is judged unsupported.
- Input the campaign letter text together with the reply deadline. Compare the current date against the deadline and check whether any reply record exists. The condition is met when "has confirmed participation" is found to lack a basis.
- Input the paper value []. Compare whether the downstream answering node proceeds with empty citations. The condition is met when the chain is not interrupted by the empty array.
All of the above is paper checking. No actual retrieval, signature, approval or settlement was performed.