What the documents look like
Wholesale stall trading produces a scattered paper trail. Purchase intake notes, handwritten sales slips, stall lease contracts and payment receipts sit alongside market price sheets, stall inventory counts, WeChat collection records and logistics consignment notes. A single transaction may leave only partial fields in each material, so item name, quantity, amount, consignment destination and payment status have to be compared across documents. Records update daily and per order. In multi-turn dialogue, a document clue raised in one turn needs to carry into the next so that order fulfilment and after-sales answers stay anchored to the same transaction.
Constraints at the multi-turn prompts step
The prompt needs to state the material scope, the answering task and the handling of missing evidence, so the model does not write in business facts that the source does not carry. History records should be checked against the context actually brought in, and a new session should be distinguished from a continued session by its input. Where a knowledge base is used across turns, check whether reference resolution and question expansion apply; short questions are not automatically covered. Model and length settings should be checked against the model's supported context and reply configuration, with long documents combined with follow-up questions tested. Reply format follows what the downstream step consumes; when json_schema is selected, the corresponding format requirement is filled in and the returned structure is verified. Retrieval returns content. Authorization, approval, database writes and external synchronization each need their own downstream implementation.
Hypothetical paper exercise: the material text reads "On 12 May 2024, stall A issued a handwritten sales slip to buyer B, item 'white radish', quantity 30 cases, amount left blank; a WeChat collection record the same day shows B transferred 2000 yuan with no order number noted; a consignment note shows dispatch to city C, consignee D." The draft sentence under review is "B has settled the full payment for this batch of white radish, and subsequent shipment can be arranged." The sentence lacks the amount field and the order-number correspondence, so it is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model shown in the configuration dialog as supporting tool calls | Compared with a model that does not support tool calls, observe whether order fields can be referenced by nodes |
| Memory turns | Set to 3 turns | Compared with 1 turn, observe whether the previous turn's document clue is carried in; compared with a larger value, observe consumption |
| Temperature | Set low | Compared with a higher value, observe whether after-sales wording stays stable |
| Reply limit | Opened and set to a medium length | Compared with a lower value, observe whether amount and status statements are truncated |
| Reply format | Kept at default | Compared with json_schema, observe whether downstream nodes can read fixed fields |
| Multimodal recognition | Image recognition enabled according to model capability | Compared with disabled, observe whether a handwritten sales slip image becomes recognizable input |
Three common errors
- Treating retrieval results as authorization. Similarity scores and
quoteListreflect returned content only. They do not confirm buyer identity, payment authority or shipment approval; those business actions need a separate system implementation. - Writing a pending state as approved. In the exercise, the amount is blank and the transfer carries no order number, so the claim is unsupported. Writing "settled" creates a business fact that the source does not contain.
- Moving the thinking budget into the citation limit. The thinking budget is a reasoning budget and is unrelated to the number of knowledge base citations; it cannot be used to infer citation scope.
How to verify the setup
- Enter the hypothetical material text above together with the draft sentence, compare "settled the full payment" against the missing amount field and the absent order-number correspondence, and observe whether the output marks the claim as unsupported. If the output directly confirms settlement, the check fails.
- Enter two consecutive follow-up turns under the same
chatId, compare whether the second answer refers to the sales slip clue raised in the first turn, and observe whether history is carried in according to the memory turns setting. If the second turn loses the earlier context entirely, review the memory turns setting. - Enter a handwritten sales slip image, compare the node input with image recognition enabled and disabled, and observe whether the image content becomes recognizable model input. If the model does not support multimodal input, the dialog shows an unsupported notice; record the gap rather than assuming recognition succeeded.