What the documents look like
Franchise fee and deposit settlement centres on funds and breach deductions, and it does not carry government filing duties. This page covers the inventory procurement and supplier records task, where the working materials are franchise fee invoices, deposit ledgers, deduction notices and contractual payment schedules. These materials also connect to purchase orders, supplier statements and goods-receipt records. The typical parties are chain finance staff, the franchise management department and franchisees. Errors concentrate in fund disputes, franchise litigation and deposit refund conflicts. The update rhythm is monthly and per deduction event; on this page that rhythm only affects how document versions are labelled, and it does not imply batch synchronisation or event-triggered automation. Materials commonly carry fields such as entity name, contract number, store number, deduction reason and payment schedule period, and retrieval must keep the object and status attached to each of those fields distinguishable.
Retrieval quality here depends on three practical facts. First, one knowledge base search node may be linked to one or more knowledge bases that use the same vector model, so a procurement collection and a settlement collection can be searched together only when their vector model identifiers agree. Second, search modes behave differently: semantic search suits natural-language questions, full-text search suits numbers, models and proper nouns, and hybrid search combines both through RRF before optional reranking. Third, a citation cap is measured in tokens rather than a fixed count of items, because chunk lengths differ greatly between question-answer and document collections. Each of these facts changes how a test question about a deposit ledger or a deduction notice should be posed and read.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that share the same vector model. Knowledge bases built on different vector models cannot be linked together. The operational step is to check the vector model identifier, exclude those that do not match, and record what was observed; a cross-model retrieval is not run to compare missed recalls. Search modes are compared against the same test question across semantic, full-text and hybrid search, and no single mode is assumed to suit all materials. The citation cap is expressed in tokens, and materials of clearly different lengths are used to see whether content is truncated before the cap is adjusted. The minimum relevance value applies only in semantic search or when result reranking is used, and labelled relevant and irrelevant samples are prepared before the threshold is tuned. Question optimisation performs coreference removal and question expansion, so the original question and the optimised question are compared to check whether the subject has been altered. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model that is enabled according to its settings. An empty result is output as an array whose length may be zero, and the downstream chain still runs.
Hypothetical paper exercise sample. Material text A reads: "Party A, XX Chain Management Co., Ltd., and Party B, a franchisee, signed a franchise contract for the XX Road store, contract number H-000; the payment schedule provides that a deposit of RMB 50,000 is to be paid on or before 31 March 2025, and the deduction reason is failure to pay on time." Material text B reads: "The statement from supplier XX Trading Co., Ltd. shows 200 units received in March 2025, purchase order number P-000." The draft sentence to be judged reads: "The franchisee's deposit has been paid, so the refund can be processed." The source text states only the payment deadline and the deduction reason; it does not state an actual payment record or a refund approval status. The draft sentence is therefore unsupported, and the evidence gap must remain open.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only knowledge bases whose vector model identifier matches | Check identifiers for consistency, exclude non-matching bases, record the actual observation |
| Search mode | Hybrid search | Compare the same test question against semantic search and full-text search separately |
| Citation cap | Starting point set to cover a complete payment schedule paragraph in tokens | Compare against a smaller value to see whether content is truncated; no official default is claimed |
| Minimum relevance | Starting point set from labelled relevant and irrelevant samples | Compare against adjacent thresholds for changes in false recalls and misses; applies only in semantic search or reranking |
| Question optimisation | Enabled for multi-turn follow-up questions that contain references | Compare the original and optimised question, and check whether the subject was altered |
| Result reranking | Used when a rerank model is available and enabled per its settings | Compare ordering against results without reranking; the rerank score is used for filtering |
Three common errors
- Treating the minimum relevance value as a permission or entity isolation control. That value applies only in semantic search or when result reranking is used, and it is a filtering threshold; it does not exclude unrelated business lines or unauthorised records. Check: submit the same test question, compare returned content at different thresholds, and observe whether any permission-related field changes; the expected observation is that none does.
- Describing RRF as a toggle switch. RRF is the fusion step for multiple recall paths inside hybrid search, not a selectable control. Check: inspect the configuration interface for a control bearing that name, record the actual observation, and leave the answer unknown where it is absent.
- Using retrieval results as a substitute for business authorisation. A hit on a franchise contract payment schedule does not mean the deposit has been paid or that a refund has been approved. Check: compare the draft sentence field by field against the source text; a field may be cited only when subject, period and status agree, and fields not stated are recorded as pending verification.
How to verify the setup
- Enter the material text from the hypothetical sample together with the draft sentence, and compare whether the draft sentence claims a payment and refund status that the source text does not state. The expected observation is that the draft sentence is judged unsupported and the evidence gap is preserved.
- Enter the same test question and run it under semantic search, full-text search and hybrid search in turn, then compare whether the returned content covers the payment schedule and the deduction reason. The expected observation is that the three modes differ, with no guarantee of a first-position hit or of complete recall.
- Use a clearly stated paper input of an empty array to check the downstream design, and observe whether the chain still runs when the cited content length is zero. The expected observation is that the downstream handles an empty citation; no claim is made that a retrieval under an invented rule necessarily returns an empty array.
At operational handover, the authorisation status of each material, action and request must be checked separately. A pending review does not demonstrate an absence of operating permission; a quotation validity period does not indicate a supply period; signing, reaching cooperation and contract effectiveness are three distinct facts. Retrieval and forms collect values, while business verification, approval, database writes and external synchronisation each require their own downstream implementation.