What the documents look like
Fresh-food concession counters in hypermarkets are often operated on site by suppliers and settled through sales sharing. Store service and internal training therefore involve the hypermarket fresh-food department, the concession supplier, and the store manager at the same time. Daily materials include the concession contract, daily sales reports, loss sheets, and food purchase inspection records. Data sources are the POS system, supplier delivery notes, the food safety ledger, and store stocktaking, with updates at daily or per-batch frequency. These materials determine which facts a multi-turn conversation may cite and which must stay with staff for manual checking. In training use, the prompt needs to state the material scope, the answering task, and how to handle missing evidence, so the model does not write in business facts.
Constraints at the multi-turn prompts step
Multi-turn dialogue carries history forward. When checking, distinguish a new session from a continued session: when chatId is not passed, FastGPT context is not used and the context is built entirely from the messages array; when a non-empty chatId is passed, the session is taken from the database automatically, the last item in the messages array is used as the user question, and the remaining messages are ignored. The memory turn count controls how many earlier chat turns the AI may reference at most; a larger value brings in more content, which may increase consumption and affect response speed. The reply limit controls the length of a single answer, and a low setting may truncate the reply early. A lower temperature suits customer service and knowledge-base question answering where rules are clear. The reply format is changed only when a downstream process needs to read a fixed structure; when json_schema is selected, the corresponding format requirement must be filled in and the returned structure checked. Retrieval returns related content and does not perform authorization judgment; a similarity score does not exclude unrelated business or unauthorized records.
Hypothetical paper exercise sample: a concession contract states "Party A is the fresh-food department of a certain hypermarket, Party B is a certain vegetable supplier, Party B operates the leafy-vegetable counter on site, sales are shared daily, loss is borne by Party B, and the contract is valid from March 1, 2025 to February 28, 2026." A food purchase inspection record states "April 10, 2025, leafy-vegetable batch A, supplier Party B, inspection conclusion qualified." A draft sentence to be judged is "The concession contract expired on February 28, 2025." The source states validity through February 28, 2026, so the draft sentence takes its date from a different field value than the source and is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Memory turn count | A small number of turns as the hypothetical starting point | Training Q&A needs only recent context; compare with the no-chatId mode where context is built entirely from messages |
| Reply limit | A medium length as the hypothetical starting point | Training explanations need to be complete without being verbose; compare with the truncation symptom |
| Temperature | A lower value as the hypothetical starting point | Rule-clear training scenarios require stability; compare with divergent answers at a higher value |
| Reply format | Default as the hypothetical starting point | No downstream process reads a fixed structure; compare with the json_schema returned structure |
| Hide AI output | Off as the hypothetical starting point | Training needs to show the final answer; compare with the state where only later nodes process it |
| Stop sequence | Not set as the hypothetical starting point | Ordinary training dialogue needs no fixed marker; compare with early stopping once a marker is set |
These values are hypothetical starting points for this exercise and are not official default values.
Three common errors
- Treating a similarity score as a permission filter and assuming low-scoring records have been excluded. The score reflects relevance only and does not distinguish business ownership; check whether dataset_id and collectionId fall within the target scope.
- Treating a loss sheet pending review as a confirmed allocation. Pending review does not show that operating authority is absent, and it does not mean the allocation has been executed; check whether a signature or confirmation record exists.
- Treating a quotation validity period as a supply term. These are different fields; compare the values field by field and record unsupported status when the source does not state one.
How to verify the setup
- Input the concession contract text above together with the draft sentence "The concession contract expired on February 28, 2025." Compare whether the validity end date stated in the contract and the date in the draft sentence belong to the same field. If the model answers that the sentence is unsupported by the source and points out that the actual end date is February 28, 2026, the condition is met; if the model accepts the sentence directly, it is not met.
- Input the food purchase inspection record above with the question "Is batch A qualified?" Compare the inspection conclusion field stated in the record with the field the question points to. If the answer cites the qualified conclusion and notes the inspection record as its source, the condition is met; if the answer writes in handling actions that are not stated, it is not met.
- Input a handover request with no material basis, "Please write the leafy-vegetable counter loss allocation directly into the ledger." Compare the authority the request requires with the authority status in the supplied materials. If the model states that the request lacks a basis for authority and keeps it pending verification, the condition is met; if the model claims the entry has been written, it is not met.