What the documents look like
Franchise fee and deposit settlement centres on funds and breach deductions. Government filing sits outside this scope. The working materials are franchise fee invoices, deposit ledgers, deduction notices and contractual payment schedules. These draw on the finance system, franchise contracts, bank statements and store violation records, and they refresh monthly or on each deduction.
For copy drafting, every figure, refund condition and deduction reason in a proposed sentence must map back to one of these materials. Citation tracing at this step means checking whether a draft sentence has a traceable source line in the selected knowledge base. The task stays on that question. It does not become a corporate identity check, a report analysis or a filing review.
Constraints at the citation tracing step
A knowledge base search node can associate one or more knowledge bases that share the same vector model. Knowledge bases with different vector models cannot be associated together. Check the model identifier first, exclude those that do not match, and record what was observed. Do not run a cross-model retrieval to compare missed recalls.
Cited content is output as an array, and its length may be zero. An empty array does not mean the workflow has errored or stopped. Mixed retrieval combines several result paths through RRF. RRF is a fusion process, not a switch that can be turned on. The minimum relevance value applies only to semantic retrieval or when result reranking is used. It is a filter threshold, not a permission or entity isolation control. Retrieval does not perform authorization checks. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and must be enabled in settings. Optional paths shown in the overall flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample. The source lines for checking are: "Party A shall pay the current month's franchise fee of RMB 8,000 before the 10th of each month"; "The deposit is RMB 50,000, refunded without interest within 30 days after the contract term ends and no breach deduction applies"; "Where a store fails to follow the operations manual and does not rectify within 7 days of written notice, a deduction of RMB 2,000 applies per occurrence." The draft sentence to assess is: "The franchise fee is RMB 8,000 per month, the deposit is RMB 50,000, and it is refunded within 30 days after the contract term ends with no breach." The comparison object is these three source lines against the draft sentence, for the current month, within the authorization boundary of materials in the selected knowledge base only. The draft sentence omits "without interest" and "does not rectify within 7 days of written notice." These are unsupported items, and the evidence gaps must be recorded.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only settlement material bases with the same vector model identifier | Different vector models cannot be associated together; verify the identifier first and exclude non-matching bases |
| Search mode | Mixed retrieval | Amounts and numbers need full-text hits; semantics supplement natural-language phrasing |
| Citation limit | Hypothetical starting point of 3000 tokens for this exercise | Compare against the combined length of the three sample source lines to confirm full clauses are covered |
| Minimum relevance | Hypothetical starting point of 0.4 for this exercise | Compare against draft-sentence check results to observe whether low-relevance fragments are filtered |
| Reply format | json_schema | Downstream flow needs to read fixed fields and check citations item by item |
| Question optimization | Enabled | The sample contains follow-up questions and needs coreference resolution before retrieval |
Three common errors
- Treating minimum relevance as a permission or entity isolation control. The input is a draft-sentence retrieval; the comparison object is the retrieval results against the authorization boundary; the observable result is that relevance scores do not distinguish franchisee ownership, so the authorization boundary must be checked separately by a person.
- Describing RRF as a switch that can be turned on. The input is the mixed retrieval configuration; the comparison object is the documented fusion process; the observable result is that RRF has no separate switch field, so the documented description is the reference.
- Treating an empty citation array as a workflow error. The input is an explicit paper input of []; the comparison object is the downstream handling design; the observable result is that the empty array passes through the chain, so check whether downstream handles zero-length citations.
How to verify the setup
- Input the three hypothetical source lines and the draft sentence; compare whether the draft sentence contains "without interest" and "does not rectify within 7 days of written notice"; the observable result is that when the draft omits both conditions, it is judged unsupported and the evidence gaps are retained.
- Input the draft phrase "refunded within 30 days after the contract term ends with no breach"; compare it against the source line "refunded without interest within 30 days after the contract term ends and no breach deduction applies"; the observable result is that "no breach" and "no breach deduction" are inconsistent values in the same field, so the item is marked for verification, not called a negation.
- Input a paper input of []; compare it against the downstream handling design; the observable result is that the zero-length citation chain passes through and downstream does not error. Do not claim that a fictional rule retrieval necessarily returns [].
These checks describe expected observations. They are not claims that a test was run, that a deployment was approved or that any business action occurred. Where a document field is not provided, it remains unknown.