What the documents look like
A brand concession inside a shopping mall operates with four data sources that do not collapse into one record: the mall POS, the brand ERP, mall membership campaign material, and the finance system. The mall collects payment centrally and settles with the brand on a commission basis, so the concession sales record and the brand-side order are separate datasets. Four document types carry the business facts across different stages: the joint-operation commission contract, the mall campaign letter, the sales receipt, and the monthly settlement statement.
For order fulfilment and after-sales answers, each cited source has to resolve to a specific document and a specific field. When that trace fails, monthly settlement differences accumulate, campaign resources shrink, and the concession can be withdrawn from the mall. The update cadence of these records does not change the tracing task itself, so it is not treated here as a functional requirement.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. A retrieval hit is not an authorization. Whether a retrieved passage is used depends on the document version and a human cross-check. The minimum relevance value takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The same material can sit in different states at different stages. Pending review does not mean the operator lacks permission. A quoted validity window does not define a supply period.
Hypothetical paper exercise sample. The annex to a joint-operation commission contract between a brand concession and a mall states: "From 1 March 2025 to 31 March 2025, the concession participates in the mall's centralised payment collection, and the commission rate is eighteen percent of sales." A mall campaign letter for the same period states: "This letter confirms the campaign window only; cost bearing and settlement method follow the joint-operation commission contract." A monthly settlement statement lists March sales and the commission amount. A mall campaign registration form sits in a "pending review" state. The draft sentence to be assessed is: "The March commission fee has been settled in full at the rate confirmed by the campaign letter."
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases holding the commission contract, campaign letter, receipts, and settlement statements | First compare the vector model identifier of each base; only bases with the same vector model can be associated together |
| Search mode | Hybrid search | Contract numbers and campaign windows are exact-match queries, while semantic questions need vector recall; compare returned citation items under each mode |
| Minimum relevance | Set to 0.5 as a hypothetical starting value for this exercise | This value takes effect only in semantic search or with reranking; compare returned item counts against 0.3 and 0.7 and adjust |
| Citation cap | Set to 2000 tokens as a hypothetical starting value for this exercise | The cap is expressed in tokens; compare against 1000 and 4000 and observe whether long settlement paragraphs are truncated |
| Result reranking | Enable ReRank | Reranking produces a 0-1 score; compare the reranked order against the unranked order item by item |
| Question optimisation | Disabled | The draft sentence has a clear subject; compare recall differences in multi-turn follow-up scenarios after disabling |
Three common errors
- Treating a retrieval hit as business authorization. Input the pending campaign registration form, compare its status field against the value "reviewed", and observe that the status remains pending review; the claim "settled in full" in the draft sentence is not supported.
- Associating knowledge bases with different vector models. Input the vector model identifiers of two bases, compare whether they match, and observe that when the identifiers differ the association is excluded and the actual observation is recorded.
- Treating a quoted validity window as a supply period. Input the campaign window wording in the campaign letter, compare its field scope against the contract settlement clauses, and observe that the letter confirms the window only while cost bearing follows the contract; the rate source in the draft sentence does not match the contract annex.
How to verify the setup
- Input the contract annex text and the draft sentence from the hypothetical sample above. Compare the rate stated in the draft sentence against the eighteen percent recorded in the annex, and observe that the two match, which satisfies the citation condition for that part. When the draft sentence claims "the rate confirmed by the campaign letter", compare against the campaign letter text and observe that the letter records no rate, so that part is not supported.
- Input the monthly settlement statement and the draft sentence. Compare the March sales and commission amounts listed in the statement against the wording "settled in full" in the draft sentence, and observe that the statement lists amounts only and records no settlement completion status, so that wording remains to be verified.
- Input the status field of the campaign registration form and the draft sentence. Compare the status value against "reviewed" and observe that the status is pending review, so the conclusion in the draft sentence that concerns this form does not hold. If the retrieval output is an empty array, check the downstream design against a citation output of length zero; do not infer that a retrieval rule must return an empty result.