What the documents look like
Franchise filing management works from the commercial franchise information management system, enterprise licences, store lists and contract ledgers. The documents handled day to day are franchise contracts, operations manuals, filing materials and information disclosure documents. Copy drafting for goods and marketing meets a defined boundary here: any store count, operating status, filing progress or contract clause that appears in proposed copy needs a corresponding statement in existing materials. Materials update on an irregular basis or by year, so a status captured at one point in time cannot be treated as a continuing basis for promotion.
Two features shape the drafting task. First, the store list and filing materials describe stores at different stages, and the stage attached to a store is part of the record rather than a rounding choice. Second, the contract ledger and disclosure documents carry their own wording, so a draft sentence that restates a number or status has to be compared against the source wording rather than against a general impression of the business.
Constraints at the multi-turn prompts step
Multi-turn dialogue separates a new session from a continued session. When chatId is empty, the system does not use the context function provided by FastGPT and builds context entirely from the messages passed in. When chatId is a non-empty string, the system retrieves the session from the database automatically, uses the last item in the messages array as the user question and ignores the remaining messages. Reply format follows the downstream requirement; ordinary question answering keeps the default, while selecting json_schema requires filling in the corresponding format requirement and checking the returned structure. Memory rounds control how many earlier rounds of chat the AI may reference at most, and a value that is too small may leave earlier information unused.
Hypothetical paper exercise sample, clearly labelled as invented for this page and not drawn from real records: the source material reads, "As of 31 December 2025, the company has 12 franchised stores open in East China; 3 of these are under renovation and not yet open. The store list in the filing materials is based on open stores." The draft sentence to be assessed reads, "East China already has 15 franchised stores in formal operation." The comparison examines the subject, object, status and quantity scope of the source wording: the source separates 12 open stores from 3 under renovation, and the draft merges them into 15 stores in formal operation, which departs from the source statement.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| AI model | Select a model whose maximum context suits the material length; check maximum context and multimodal capability | The sample material is text, so multimodal recognition is not required |
| Memory rounds | Set the number of earlier chat rounds to retain according to the dialogue need | A follow-up question about store counts needs the earlier statement |
| Reply limit | Adjust the reply length limit according to the copy length needed | The draft sentence is short, so a high limit is unnecessary |
| Temperature | Copy drafting may be raised somewhat; rule-checking scenarios are better kept low | The sample involves factual comparison and is better kept stable |
| Reply format | Ordinary question answering keeps the default; a fixed structure requires json_schema with the format requirement filled in | The sample is a single-sentence assessment, so the default format is sufficient |
| Hide AI output | Enable when an intermediate result is passed to a later node | The sample outputs the assessment directly, so this is not needed |
The starting values above are implementation checks for this page, not documented defaults. Where a value is chosen, the reason is the nature of the sample task and the comparison is against the source material wording.
Three common errors
- Counting stores recorded as under renovation and not yet open as stores in formal operation, so the draft sentence departs from the source status. The input is the source material and the draft sentence; the comparison is between the status attached to each store group in the source and the status asserted in the draft; the observable result is a merged count that does not match the separated groups.
- Treating every item in the messages array as participating in context when chatId is non-empty. The input is a request with a non-empty chatId and several messages; the comparison is between the messages actually used and the documented rule that only the last item serves as the user question; the observable result is a reply that reflects earlier messages when the documented behaviour would ignore them.
- Treating a retrieval relevance score as a business filter that excludes unfiled or unauthorised stores. The input is a retrieval request over the store list; the comparison is between the returned records and the records actually stated in the material; the observable result is a claim about authorisation or filing status that the score does not carry.
How to verify the setup
- Enter the hypothetical source material and the draft sentence given above. Compare the store count, operating status and scope in the draft against the source wording. If the draft states 15 stores in formal operation by merging the 12 open stores with the 3 under renovation, the draft is judged inconsistent with the material. A draft that keeps 12 open and 3 under renovation separate is consistent with the source.
- Enter a follow-up question under the same chatId and compare the context rounds actually carried into the reply against the memory rounds setting. If the follow-up can refer to the store count stated earlier, the history carried in matches the setting. If the follow-up cannot refer to it, the carried context does not match the setting and the reason remains to be checked.
- Enter a retrieval request covering the store list and observe whether returned records include stores not stated in the material. If the returned records contain only stores recorded in the material, the retrieval scope matches the material. If stores outside the material appear, record them as pending verification rather than as an authorisation or filing result.
A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the material does not state a field, the field remains unknown. The relevance score is not a category, entity or permission filter, and retrieval is not authorisation. Forms collect values; business verification, approval, database writes and external synchronisation require their stated downstream implementation.