What the documents look like
Platform store ratings and experience score maintenance centers on consumer feedback that arrives after fulfillment is complete. Compared with promotional campaign registration, this work does not directly determine campaign pricing. The primary data comes from Douyin store experience scores, Tmall service indicators, JD after-sales reviews, and customer service systems. Daily documents include negative-review handling records, experience score detail sheets, service diagnostic reports, and rectification ledgers. In store service and internal training, customer service supervisors and quality inspection specialists use multi-turn prompts to link negative-review attribution, supplementary evidence, and rectification tracking into a checkable process. Data updates daily or weekly, but the feature on this page does not depend on batch synchronization or real-time triggers.
Constraints at the multi-turn prompts step
Multi-turn dialogue depends on history records being carried into the actual context. When a new session does not pass a chatId, FastGPT does not use its context feature and builds context entirely from the passed messages. When a non-empty chatId is passed, the session is taken automatically from the database, the last message in the messages array is used as the user question, and the remaining messages are ignored. In store training where agents practice consecutive follow-up questions, the actual carried context must be checked, distinguishing new-session input from continued-session input.
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. For question optimization, when a knowledge base is used in continuous dialogue, check reference resolution and question expansion; do not assume all short questions are suitable. For model and length, check the context and reply settings the model supports, and test long-document and follow-up combinations. Reply format is chosen according to downstream handling; when json_schema is selected, fill in the corresponding format requirements and verify the returned structure. Semantic recall and relevance scores do not handle permission or entity isolation. The requirement that a retrieval library use the same vector model cannot be extended to mean external systems must also use the same model.
Configuration and implementation checks
The following configuration concerns the hypothetical drill sample on this page. The sample is paper material and does not involve a real customer or legal identity number.
Hypothetical drill sample: negative-review handling record excerpt — "Order A was signed for on March 5; the consumer submitted a negative review on March 8, reporting that the outer packaging was damaged. Customer service contacted the consumer on March 9; the consumer did not provide an unboxing video. The rectification ledger shows the store must complete packaging reinforcement rectification in March; the current status is pending review." One draft sentence awaiting judgment: "This negative review has been confirmed as caused by logistics damage, and the store does not need rectification."
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| AI model | Select a model that supports multimodal recognition | Negative-review records may contain image evidence; compare with the model capability description |
| Memory turns | Keep a small number of history turns | Customer service scenarios usually keep a small number of history turns; compare with the default value |
| Temperature | Low | Customer service and knowledge base Q&A suggest a low value; compare with the default value |
| Reply format | Default | Ordinary customer service Q&A keeps the default; compare with a json_schema requirement |
| Hide AI output | Off | Training scenarios need to observe intermediate results; compare with behavior when on |
| Thinking configuration | Default | Use the model default when there is no special requirement; compare with each level |
Three common errors
- Treating retrieval results as an authorization basis. In the hypothetical drill sample, the rectification ledger status is "pending review." If the draft sentence says "the store does not need rectification," the input is the ledger status field, the comparison is between "pending review" and "does not need rectification," and the two are inconsistent. The observable result is that the draft sentence contradicts the ledger status field.
- Treating an offered option as executed. If a handover request says "the packaging reinforcement plan has been selected and rectification completed," the input is the option record in the ledger, the comparison is between "selected" and "completed," and selecting an option does not equal executing rectification. The observable result is that the completion claim in the request lacks execution evidence.
- Treating related-stage evidence as a negative conclusion. In the hypothetical drill sample, the consumer did not provide an unboxing video; this only shows that the evidence is missing, and cannot be used to judge that the negative-review cause does not hold. The input is the missing-evidence record, the comparison is between "video not provided" and "logistics damage confirmed," and the two are not mutually exclusive values of the same field. The observable result is that the draft sentence turns unsupported content into confirmed content.
How to verify the setup
- Input the original negative-review handling record text and the draft sentence from the hypothetical drill sample. Compare "confirmed as caused by logistics damage" in the draft sentence with "the consumer did not provide an unboxing video" in the original. If the original does not state a determination of the damage cause, the draft sentence is unsupported and the evidence gap must be recorded.
- Input the "pending review" status in the rectification ledger and "does not need rectification" in the handover request. Compare the values of the same field. If the two are inconsistent, the handover request contradicts the ledger status and must be returned for a supplementary review conclusion.
- Input the customer service system session record and the chatId passing situation. Compare the number of context turns actually carried with the memory turns setting. If history turns in a continued session exceed the setting, check the session record corresponding to the chatId to confirm whether the carried content matches expectations.