What the documents look like
This page covers copy drafting for four document types in retail checkout and loyalty promotion: promotional price tags, membership points rules, the checkout daily report, and abnormal refund slips. Promotional prices and membership benefits sit close to price complaints and cash shortfalls, so any drafted sentence that states a price, a points multiplier, an applicable store scope, or a validity period must match the material it draws from. The checkout daily report and abnormal refund slips update by shift or by day, and drafting works from the currently recorded figures for that period, so wording from a previous shift or a previous day cannot be carried forward as if it described the current one.
Multi-turn dialogue serves to complete the product, campaign scope, and benefit conditions across several follow-up questions. The prompt constrains the model to write sentences only from the supplied material, and to leave gaps for verification when evidence is absent.
Constraints at the multi-turn prompts step
History determines which prior text the model can see. When chatId is passed, FastGPT retrieves the conversation from its database automatically and uses the last element of the messages array as the user question, ignoring the remaining messages. When chatId is not passed, FastGPT context is not used and the context is built entirely from the messages supplied in the request. The prompt must state the material scope and how to handle missing evidence, so the model does not invent prices, points, or refund facts. Reply format follows what the downstream step needs to read; choosing json_schema requires filling in the corresponding format requirement and checking the returned structure. Model context length and reply cap affect the combination of long material plus follow-up questions, and a memory turn count that is too small loses earlier text.
Hypothetical drill sample: the source material reads "This promotion covers the A store's personal care category, and the promotional price follows the price tag system record for the current day. The membership points rules state that this category accrues base points, and the doubling multiplier during the campaign follows the current CRM announcement. An abnormal refund slip must be submitted by the checkout supervisor within the same shift." The proposed draft sentence reads "A store personal care items earn doubled points during the campaign, and the refund arrives the same day." Assessment: the points multiplier points to the current CRM announcement in the material, and the draft sentence does not supply that announcement's content, so it is unsupported; the refund arrival time is not stated in the material, so it remains to be verified.
Configuration and implementation checks
| Configuration item | Value or check item on this page | Relation to the sample |
|---|---|---|
| AI model | Select a model that supports the needed context length and multimodal capability | Check maximum context when long material combines with follow-ups |
| Memory turns | Hypothetical trial starting point of retaining a small number of turns, because customer-service style question answering usually needs only a little history; compare against an input that retains no history | Check whether earlier text is still present during follow-up |
| Reply cap | Hypothetical trial starting point of medium length, because a complete explanation must be generated; compare against a truncation case | Check whether the draft sentence is cut off early |
| Temperature | Lower, because the copy must stay close to the material | Reduces the chance the rewrite drifts from the material |
| Reply format | Default; when a fixed structure is needed, choose json_schema and fill in the format requirement | Check the returned structure against downstream reading |
| Multimodal recognition | Enable only when the model supports it and the input contains images | Product image and poster recognition depends on model capability |
Three common errors
- Writing a points multiplier as a definite value when the material does not state it. The draft sentence says "doubled points," while the material points to the current CRM announcement without giving a multiplier. Check: input the material text and the draft sentence, compare whether the multiplier has support from the current announcement, and record it as unsupported when the announcement is not provided.
- Treating the refund arrival time as known. The material states only that an abnormal refund slip is submitted by the checkout supervisor within the shift, and does not state an arrival time. Check: input the material and the draft sentence, compare whether "arrives the same day" has a corresponding field, and mark it as pending verification when there is none.
- Using a similarity score in place of an authorization or entity judgment. A retrieval score does not constitute a store or member permission filter. Check: input a query carrying a store identifier, compare whether the returned records' store identifier matches the target scope, exclude those that do not match, and record the observation.
How to verify the setup
- History check: with the same chatId, send the material sentence first and then a follow-up, and observe whether the model references the earlier material; compare against a request that does not pass chatId, to confirm the difference in context source. The expectation is that passing chatId uses the database conversation, and not passing it uses only the supplied messages.
- Draft sentence check: input the material text above and the draft sentence, compare the two items of points multiplier and refund timing, and expect to obtain unsupported and pending verification, with no definite numeric value appearing.
- Empty-result handling check: use the explicit paper input [] to check how the downstream design handles empty references, and observe whether the downstream proceeds as designed; a fabricated rule retrieval cannot be claimed to return [] with certainty. Libraries across different vector models cannot be associated together, so check identifiers only, exclude those that do not match, and record the actual observation.