What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, without franchise contracts governing the relationship. Store service and internal training conversations therefore draw on internal systems rather than franchise reporting. The daily document set consists of daily sales reports, shift schedules, display photos, and inventory transfer orders, with data updated daily and weekly. Three aspects of this material shape multi-turn prompts. First, the reference scope must be bounded, because training questions often cite same-day report values and transfer order status. Second, history carry-over must be controlled, because a continued session pulls context from the database while a new session does not. Third, the reply format must be selected according to what the next node can accept. When the prompt does not state the document boundary, the model can supply inventory or authorization facts that the source material never contained.
Constraints at the multi-turn prompts step
The prompt must state which documents and which time range may be cited, and must state that a question outside that range receives a statement of inability to answer. For history, a request without a chatId does not use the context feature and builds context entirely from the supplied messages array; a request with a non-empty chatId takes the session from the database and uses only the last item of the messages array as the user question, ignoring the remaining messages. Where knowledge-base continuity is needed, check whether reference resolution and question expansion apply, and treat short questions as not automatically eligible. For model and length, check the context size and reply ceiling that the selected model supports, and test long documents combined with follow-up questions. Reply format follows the downstream consumer; when json_schema is selected, the corresponding format requirement is filled in and the returned structure is verified. Retrieval and similarity scores do not perform permission isolation, and a retrieval action does not itself constitute authorization.
Configuration and implementation checks
The following hypothetical drill sample is used throughout this page. A brand-owned store shift schedule dated 10 June 2025 reads: "Morning shift 3 people, evening shift 2 people, store manager off." An inventory transfer order reads: "Transfer order number DB-0610-07, status is pending review, source warehouse: Store A, destination warehouse: Store B, quantity 12 items." A draft sentence under review reads: "Store B has received the 12 items transferred from Store A." The source text records only that the transfer order is pending review, and records no receipt or effectiveness. The draft sentence exceeds the source in both subject and status.
| Configuration item | Setting applicable to this page | Check method |
|---|---|---|
| Memory turns | Hypothetical trial starting value 3 | Compare with 0 and 6, observe whether earlier turns enter the reply |
| Reply ceiling | Hypothetical trial starting value 800 | Compare with 400 and 1600, observe truncation and completeness |
| Temperature | Hypothetical trial starting value 0.3 | Compare with 0.7, observe reply stability |
| Reply format | Default; select json_schema when a downstream node needs a fixed structure | Inspect whether the returned structure matches the format requirement |
| Multimodal recognition | Enable only when the model supports it and display photos must be read | Confirm whether the dialog shows the corresponding option |
| Hide AI output | Enable when a downstream node must consume that output | Check whether the following node reads that output |
The memory-turn and reply-ceiling values are trial starting points chosen to sit between a smaller and a larger comparison value, not documented defaults. Temperature is chosen low because training answers should be stable, and is compared against a higher value to observe variation.
Three common errors
- Writing a pending-review transfer order as completed receipt. The input is the transfer order text above together with the draft sentence. The comparison is between "pending review" and "has received." The draft sentence exceeds the source range and is unsupported.
- Treating a similarity score as a permission filter. The input is the score field returned by retrieval. The comparison is between that score and business authorization. The score does not indicate that a record has been authorized.
- Treating a thinking budget as a citation token ceiling. The input is a thinking configuration level. The comparison is between that level and the citation length setting. These are different parameters.
How to verify the setup
- Input the shift schedule text above with the question "How many people are on the evening shift on 10 June." Compare the reply against "evening shift 2 people." A reply that cites only that figure satisfies the condition; any other headcount does not.
- Input the transfer order text above with the draft sentence. Compare "pending review" against "has received." The condition is satisfied when the draft sentence is marked unsupported.
- Input a paper empty result
[]explicitly. Check the downstream node's design for handling an empty reference. The condition is satisfied when the output does not fabricate citation entries.
These checks state expected observations. No test, deployment, approval, or refund has been performed. Where the source text is silent, the field remains unknown.