What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and settles against a concession fee. Sales receipts are produced by the mall POS, while the brand ERP and the finance system each keep their own records, and the monthly reconciliation statement is compiled in the following month. The same sale therefore appears in the mall POS, the brand ERP and the sales receipt, and the amount and concession fields across those records must correspond item by item.
The four materials used for tracing are the joint-operation concession contract, the mall activity letter, the sales receipt and the monthly reconciliation statement. Their subjects, periods and field definitions differ, so the version and applicable period must be confirmed before any comparison. Data updates follow a daily and monthly rhythm, and this page covers only citation tracing for product information and selection questions.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model for vector search. Knowledge bases built on different vector models cannot be associated together; the identifier must be checked first and any non-matching base excluded. The output is an array of cited content whose length may be zero. An empty array means the output path still runs through, and it does not mean the workflow raised an error and stopped.
The minimum relevance score takes effect only under semantic retrieval or when result re-ranking is used. It is a filtering threshold and does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems. RRF is the fusion process applied to mixed retrieval results and is not a switch that can be turned on. Question optimization, image retrieval and result re-ranking are optional paths whose execution depends on the input and on model capability, and they are not fixed steps in every request.
Hypothetical drill sample: the source material reads "Party A, the mall, and Party B, the brand, signed a joint-operation concession contract, contract number H-2024-031, valid from 1 January 2024 to 31 December 2024, with the concession rate calculated at 18% of the month's tax-inclusive sales," "the May 2024 sales receipt summary shows tax-inclusive sales of RMB 120,000," and "the May 2024 monthly reconciliation statement records concession fees payable of RMB 21,600, with status pending confirmation by the brand." The draft sentence to be judged is "The concession fee for this counter in May 2024 has been settled." The source only shows the statement as pending confirmation and records no payment or settlement, so the draft sentence is not supported and the evidence gap must be recorded.
Configuration and implementation checks
| Configuration item | Assumed trial starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the product information base and the contract base that share the same vector model | Compare identifiers against a base on another vector model; if the identifiers differ, leave it unselected |
| Search mode | Hybrid retrieval | Compare with semantic retrieval alone for hits on numbers and contract numbers |
| Minimum relevance score | Starting point set to 0.4 | Compare returned counts at 0.3 and 0.5, and observe whether number-type questions are filtered out |
| Citation limit | Starting point set to 2000 tokens | Compare with 1500 and 3000 to see whether returned content covers the contract subject and period |
| Result re-ranking | Enabled as configured | Compare ordering with re-ranking off, and confirm the re-ranking score is used for filtering |
| Question optimization | Enabled for multi-turn follow-up questions | Compare retrieval results with it off to observe how references are completed |
These values are starting points for this drill and are not official defaults. The thinking budget belongs to model thinking configuration and is not used as the citation token limit.
Three common errors
- Treating the minimum relevance score as a permission or entity isolation mechanism. The input is a question spanning multiple counters; compare whether returned content contains only records within the authorized scope, and observe that results can still return material from other counters, which shows the threshold does not perform isolation.
- Treating an empty citation array as a workflow failure. The input is a query with no matching material; compare the node output with the downstream design, and observe that the output is an array of length zero while the path continues, which shows an empty array is not an error.
- Writing a pending reconciliation statement as settled. The input is the statement text above together with the draft sentence; compare the status field with the phrase "has been settled," and observe that the source records no settlement, so the sentence is not supported.
How to verify the setup
- Input the contract text above with the question "What is the concession rate for H-2024-031?" Compare whether the cited content contains the contract number, the validity period and the 18% concession rate, and observe that the cited item matches the contract text to satisfy the condition.
- Input the sales receipt summary and the reconciliation statement text above, compare the draft sentence "The concession fee has been settled" with the statement's status field, and observe that the status is pending confirmation by the brand, so the sentence is not supported and the evidence gap is recorded.
- Input a paper-level empty array [] to check the downstream handling design, compare whether the downstream node processes it along the empty-citation branch, and observe that the path runs through without an error to satisfy the condition for checking empty-result handling. All of the above are paper checks and do not represent an actual deployment or execution.