What the documents look like
Franchise fee and deposit settlement centers on fund receipts, payments and breach deductions. It does not carry the government filing duty that belongs to franchise registration administration. The main data comes from the finance system, franchise contracts, bank statements and store violation records. Day-to-day documents include franchise fee invoices, deposit ledgers, deduction notices and contract payment schedules.
Inventory procurement and supplier records form a separate money stream inside this category. They cover purchase orders, supplier statements of account, goods receipt notes and payment schedules. These documents differ from the deposit ledger in fund scope, even though both may pass through the same settlement team. Document parsing and chunking works on the source structure of these materials: amounts, payment terms, deduction reasons, supplier names and document numbers are scattered across tables, clauses and footnotes. The brief gives monthly updates and updates at each deduction. This page does not derive batch synchronization or real-time triggers from that frequency.
The most time-consuming steps in this category are checking contract payment terms, confirming deduction reasons and issuing invoices. Chunking decisions made here affect whether those steps can retrieve the right clause alongside the right amount.
Constraints at the document chunking step
Chunking must preserve the document identity and version that allow a retrieved fragment to point back to its source. A conditional clause and the amount or deadline it constrains must stay in the same data group. If a statement's main amount sits in one chunk and its footnote sits in another, retrieval can return the amount without the condition that limits it.
FastGPT separates raw data from retrieval indexes. One data group can map to multiple vectors, and the indexes field records the corresponding vector IDs as an array. Retrieval recalls vectors first, then looks up the original content in MongoDB by vector ID and merges results that belong to the same data group, taking the highest score. Because one data group can hold several indexes, chunk granularity directly shapes recall range.
Search modes carry different strengths. Semantic search relies on vector distance. Full-text search suits numbers, models and proper nouns. Hybrid search merges multiple recall paths through RRF. ReRank produces a 0-1 score for filtering. The minimum relevance value is a filter threshold that applies only in semantic search or with result re-ranking. It does not perform permission or entity isolation, and retrieval itself is not authorization.
Image retrieval requires image input plus the matching model capability. Re-ranking requires a model and applies according to configuration. The paths drawn in the overall retrieval diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: a franchise settlement clerk prepares to hand a supplier statement of account to a reviewer. The source text reads: "Supplier A issued a statement of account on 1 March 2025, stating that RMB 120,000 in procurement payables was due as of 28 February 2025, with payment due 30 days after reconciliation. A note below the statement reads: this amount does not include return deductions not yet confirmed." The proposed handover request reads: "Please arrange payment according to the statement amount." The point to determine is whether the statement amount has already deducted the return deduction, and whether the payment request holds authorization.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Comparison and explanation |
|---|---|---|
| Reference upper limit | Hypothetical starting point 3000 tokens | Compare against the actual token volume of the chunked statement, so the conditional note is not cut off |
| Minimum relevance | Hypothetical starting point 0.4 | Compare against semantic or re-rank scores; acts as a filter threshold only |
| Search mode | Hybrid search | Amounts and numbers use the full-text path, clause descriptions use the semantic path, merged through RRF |
| Chunk granularity | Segment by clause and table row | Compare against source structure, confirming the note stays in the same data group as the main amount |
| Question optimization | Off for this exercise | Compare against a single-turn explicit question; enabling it adds a model call and latency |
| Post-change review | Check training and index status after changes | Compare against upload success status; upload success does not mean retrieval has updated |
These values are hypothetical starting points for this exercise, not official defaults. The reason for choosing them is to allow comparison against the sample chunk length and scores.
Three common errors
- Splitting the statement's main amount and its "does not include return deduction" note into different chunks, so retrieval hits the amount but loses the limiting condition. Check: input the source text above, compare whether the note and the main amount land in the same data group after chunking, and observe whether the retrieval result carries the limiting phrase together with the amount.
- Treating a retrieval hit as payment authorization, using the similarity score to infer the procurement payment has been approved. Check: input the proposed handover request, compare the retrieval result against the authorization status field, observe whether an approval or authorization record appears, and record the gap as pending verification when it does not.
- Treating upload success as an updated index and searching directly against the new statement amount. Check: input the changed statement version, compare upload status against actual training and index status, and observe whether retrieval still returns the old amount.
How to verify the setup
- Input the statement source text above together with the proposed handover request. Compare whether the main amount, payment deadline and return deduction note stay in the same data group in the chunked result. Observe whether the retrieval result carries all three items together. A split conditional clause means the condition is not met.
- Input two queries, "Supplier A statement 120000" and "how is the return deduction handled". Compare the recall results from the full-text path and the semantic path. Observe whether the RRF merge covers both the number and the clause description. When only one path hits, record the actual observation.
- Input the changed statement version. Compare upload status against training and index status. Observe whether retrieval returns the new amount. If it returns the old amount, record that the index has not updated, without inferring a specific error cause. For an empty-result case, a clearly paper-based input
[]can be used to check the downstream design, without claiming that a fabricated rule search necessarily returns empty. Libraries built on different vector models cannot be associated together; check the identifiers, exclude those that do not match, and record the actual observation.
This page addresses only franchise fee and deposit settlement / inventory procurement and supplier records / document parsing and chunking. Whether materials, operations and requests sit in an authorized state during operational handover is a separate question. Retrieval and forms do not constitute authorization to execute.