What the documents look like
Franchise filing management works with the commercial franchise information management system, enterprise licences, store lists and contract ledgers. The daily documents are franchise contracts, operations manuals, filing materials and information disclosure documents. In store service and internal training, a multi-turn conversation has to carry contract clauses, manual steps and disclosure wording as separate bodies of material. Statements about the two-stores-one-year condition, the filing entity and the disclosure point come from specific documents, and anything those documents do not state stays unknown. Material updates are irregular and yearly, so the configuration does not need extra machinery for a sync frequency.
Constraints at the multi-turn prompts step
Multi-turn context comes from the history associated with chatId. The last item in the messages array is used as the user question and the remaining messages are ignored. When chatId is not passed, the platform does not use its stored conversation and the context is built entirely from the supplied messages. Prompts need to state the material scope, the answering task and how to handle missing evidence, so the model does not write in business facts. Question optimisation, when used for continuous knowledge base dialogue, is checked for pronoun resolution and question expansion, and whether a short question is suitable depends on the actual configuration. Model and length are checked against the maximum context and the reply cap, and long manuals combined with follow-up questions need measurement. Reply format follows what the downstream step consumes; choosing json_schema requires filling in the corresponding format requirement and checking the returned structure. Retrieval and similarity only affect recall ordering and do not perform permission or entity isolation.
Hypothetical drill sample: the source text to be checked reads "Party A is the franchisor and Party B is the franchisee. The signing date of this contract is 1 March 2024, and the filing material list sets out the franchise contract and the operations manual. The information disclosure document states that the disclosure point is thirty days before signing." A draft sentence to be judged reads "This contract took effect on 1 March 2024 and the filing has been completed." The words "took effect" and "filing has been completed" have no corresponding record in the source text, so the draft sentence is unsupported and the evidence gap is preserved rather than rewritten as a negative conclusion.
Configuration and implementation checks
The table lists official documented settings or implementation checks relevant to this feature, around the sample above. The values are starting points for this hypothetical trial, with the reason for choosing them and the object of comparison; they are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| AI model | Choose a model that supports the required context length | Compare with the combined length of the contract, manual and disclosure document in the sample, and confirm the maximum context covers it |
| Memory turns | Keep a small number of history turns first | Compare with the store service follow-up scenario, and observe whether earlier turns affect the answer |
| Reply cap | Turn on and set a medium length | Compare with the completeness of the disclosure wording, and avoid early truncation |
| Temperature | Set low | Compare with the stability requirement of filing wording, and keep answers consistent |
| Reply format | Keep the default | Compare with the downstream step, and change it only when a fixed structure is needed |
| Thinking configuration | Use the default | Compare with question complexity, and do not raise the level for simple question answering |
Three common errors
- Treating retrieval similarity as a permission or entity filter. The input is the disclosure document in the sample together with records of an unrelated franchisee; the comparison object is whether the recall result contains only the authorised scope; the observable result is that a high or low similarity has no bearing on authorisation, and permission checking needs a separate implementation.
- Treating a pending review as approved, or a signature as taking effect. The input is the signing date in the sample; the comparison object is whether the source text records a condition for taking effect; the observable result is that the source text only gives a signing date, so taking effect is unsupported.
- Treating a request without
chatIdas still using platform context. The input is a request withoutchatId; the comparison object is whether themessagesare used in full; the observable result is that the platform does not read the stored conversation and the context is built entirely from the suppliedmessages.
How to verify the setup
- Input the sample source text and the draft sentence "This contract took effect on 1 March 2024 and the filing has been completed." Compare whether the draft sentence can find a record of taking effect and of filing completion in the source text. The observable result is that the source text only gives a signing date and a material list, so the judgement is unsupported and the gap is preserved.
- Input a short follow-up containing a pronoun together with
chatId. Compare whether the pronoun after question optimisation points to the correct object. The observable result is that the condition is met only when the optimised result matches the object in the sample; when question optimisation is not configured, record it as not enabled. - Input an empty reference list
[]together with a downstream design. Compare whether the downstream step handles it through an empty-result branch. The observable result is that the condition is met only when the downstream step gives an explicit empty-result handling; do not claim that a fictional rule search necessarily returns[]. For libraries on different model versions, check the identifiers first and exclude those that do not match, record the actual observation, and do not run a cross-model retrieval to compare missed recall.