What the documents look like
Franchise fee and deposit settlement handles money and breach deductions. It does not handle government filing. The working materials come from finance systems, franchise contracts, bank statements and store violation records. The everyday documents are franchise fee invoices, deposit ledgers, deduction notices and contract payment schedules.
Cross-border and multilingual work changes what retrieval has to match. A single deposit refund request can carry a Chinese deduction notice and an English payment schedule at the same time. Retrieval therefore has to combine cross-language semantic understanding with exact identifier matching. Materials update monthly and on each deduction. This page covers retrieval and recall only. It does not infer batch synchronization, full synchronization or real-time triggers.
Two features of this category shape the task. First, the subject is funds and deductions, so the decisive evidence is an amount, a due date, a deduction reason and a settlement status. Second, the same case can be split across languages and document types, so a query in one language must reach a passage written in another.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The correct action is to check the model identifier and exclude those that do not match. Do not run a cross-model search in order to compare missed recalls.
Semantic retrieval supports cross-language understanding. Full-text retrieval suits numbers, models and proper nouns. Hybrid retrieval merges the two paths through RRF. Result reranking uses a ReRank model for a second ordering pass. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. Retrieval results do not carry permission or entity isolation duties. The same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems.
Hypothetical paper exercise sample. Material text one: "March 2025 deduction notice: Franchisee A is deducted RMB 8,000 from the deposit for failing to complete the quarterly training under Article 7 of the contract; the deduction reason has been confirmed by the franchise management department." Material text two: "Payment schedule is Franchisee A's deposit balance is to be arranged for refund after review in April 2025, and the refund is conditional on the deduction being settled." Draft sentence to be assessed: "Franchisee A's deposit was refunded in April 2025." The source text states only that a refund is to be arranged after the April review, conditional on the deduction being settled. It does not state that the refund was executed.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the Chinese contract base and the English payment schedule base that use the same vector model | Check that the vector model identifiers match, then associate; exclude those that do not match |
| Search mode | Use hybrid retrieval for deduction-reason questions and full-text retrieval for identifier questions | Compare semantic, full-text and hybrid results; do not assume one mode fits all materials |
| Reference limit | Set the starting point to a token amount that covers one deduction notice plus one payment schedule passage | Test with materials of different lengths for content completeness, then compare against a larger or smaller value |
| Minimum relevance | Set the starting point to a value that keeps the deduction-reason passage | Label relevant and irrelevant samples first, then compare results before and after filtering |
| Question optimization | Enable for multi-turn follow-up questions; disable for already explicit identifier questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a ReRank model is available and candidates are numerous | Compare scores and ordering before and after reranking and observe whether results fit the question more closely |
Three common errors
- Treating minimum relevance as a permission or entity filter. Minimum relevance only filters low-relevance results. It does not exclude unrelated business records or unauthorized records. Check: input the same question, compare returned content before and after filtering, confirm whether unauthorized records are absent only because their similarity is low, and record that retrieval does not carry authorization duties.
- Treating a pending refund request as a completed refund. The source text states only "to be arranged for refund" and "conditional on the deduction being settled," and does not state that the refund was executed. Check: compare the draft sentence with the source text, confirm that the source lacks evidence of the subject, date and amount of a completed refund, and mark the draft sentence as unsupported.
- Associating knowledge bases that use different vector models. Check: verify the vector model identifier of each base, exclude those that do not match, and record the actual observation. Do not infer the runtime result of associating them.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence, compare "to be arranged for refund" with "was refunded," observe whether the draft sentence is marked unsupported, and preserve the evidence gap about refund execution.
- Input "failing to complete the quarterly training under Article 7 of the contract" from the deduction notice as the query, compare semantic retrieval results with full-text retrieval results, observe whether the article number and the reason are recalled, and record a miss as a miss with manual review retained.
- Input an explicit empty reference content array, and check whether the downstream design follows the reference path whose length can be 0. Do not claim that a search under a fictional rule necessarily returns an empty array.