What the documents look like
For platform shop ratings and experience score maintenance, product and marketing copy drafting draws on three families of material: negative-review handling records, experience score detail, and service diagnostic reports together with rectification ledgers. These materials are keyed to a shop, an order, or an after-sales ticket. The object is consumer feedback following one fulfillment, and status fields record the handling stage and whether supporting evidence has been supplied. Records update daily or weekly, and a single record's valid scope is usually limited to a specific order and after-sales period.
Experience score detail and service diagnostic reports supply metric definitions. Negative-review handling records and rectification ledgers supply factual descriptions. Copy drafting extracts concrete statements that these materials can support. Selling points, commitments, or outcomes that the materials do not state cannot be written into a draft.
The constraint at the multi-turn prompts step is that the model must separate material facts from draft wording, and on follow-up must hold to the same subject, the same period, and the same field. Memory rounds determine how much prior conversation is referenced, the reply cap determines single-output length, and temperature affects wording stability. The prompt needs to state the material scope, the answering task, and how to behave when evidence is missing, so the model does not fill in business facts.
Constraints at the multi-turn prompts step
A status description in a negative-review handling record and a metric value in experience score detail belong to different fields and cannot be derived from each other. A rectification ledger records rectification progress; a service diagnostic report records diagnostic conclusions. These cannot be merged into one fact.
Across turns, the order discussed in the previous turn and the order asked about in the next turn must stay the same subject. If the material covers only one after-sales stage, a draft sentence cannot state that the order has completed all handling. Metric values in experience score detail carry a statistical period, and a copy reference cannot compare across periods.
The prompt needs to state that information the materials do not contain stays missing, and that business facts are not filled in. Memory rounds affect how much conversation the model can reference, and do not change the valid scope of the materials themselves. Reply format follows the downstream consumer; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure.
Configuration and implementation checks
The following items concern this page's primary task and list only documented settings or implementation checks related to multi-turn prompts.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Memory rounds | Keep a small number of history rounds | Customer service and knowledge-base Q&A applications typically keep a small number of history rounds; compare whether the model cites earlier conversation beyond the material scope at different round counts |
| Temperature | Low | Copy drafting needs stable wording, consistent with customer service and knowledge-base Q&A; compare whether draft sentences at the same material produce statements the material does not state under different temperatures |
| Reply cap | Adjust to draft sentence length needs | Once opened, a slider limits reply length; compare whether a low setting truncates draft sentences, and whether a high setting generates explanations beyond the task scope |
| Reply format | Keep default | Ordinary copy drafting does not need a fixed structure; if a downstream process reads fixed-format content, select json_schema and fill in the format requirement |
| Multimodal recognition | Enable as needed | If materials contain product images or screenshots, enable image recognition; when the model does not support it, the dialog does not show the option, and a model supporting multimodal input is needed |
| Hide AI output | Per debugging need | Once opened, generated content is not shown directly and can still pass through AI reply output to later nodes; during debugging it can be temporarily closed to observe intermediate content |
These values are hypothetical trial starting points for this exercise, not official defaults. The actual choice is determined by comparing against material scope, draft sentence length, and the downstream consumer.
Three common errors
- Writing an experience score detail metric value directly into a draft sentence without checking whether the metric's statistical period matches the period stated in the copy. The input is an experience score detail value for one period and a draft sentence; compare the two period fields. If the period stated in the draft sentence differs from the metric's statistical period, the statement is unsupported.
- In a multi-turn follow-up, applying the order facts discussed in the previous turn to a different order asked about in the current turn. The input is a two-turn conversation record and a draft sentence; compare the subject identifiers in the two turns. If the subjects differ while the draft sentence carries over the previous turn's facts, this is an evidence gap.
- The prompt does not state how to behave when evidence is missing, and the model fills in business facts where the material is silent. The input is the prompt text and a draft sentence; compare the concrete statements in the draft sentence against the material text. If a statement has no corresponding content in the material, it is unsupported.
How to verify the setup
Hypothetical exercise sample: the material text is "Negative-review handling record: order A, after-sales stage is contacted consumer to supplement evidence, evidence not yet received." "Experience score detail: shop B, statistical period is March 2025, service metric value is 4.6." The draft sentence to be judged is "This shop's service experience score is 4.6, and all after-sales issues have been fully handled."
- Input the material text above and the draft sentence. Compare "service experience score is 4.6" against the experience score detail metric value and statistical period, and compare "all after-sales issues have been fully handled" against the after-sales stage in the negative-review handling record. If the draft sentence does not state the statistical period, or if "fully handled" is inconsistent with "evidence not yet received," the statement is unsupported.
- Input the material text above and the prompt text. Compare whether the prompt states that information absent from the materials stays missing. If the prompt does not state this, and the draft sentence contains a statement the material does not state, the prompt needs the missing-evidence handling added.
- Input the material text above and a follow-up question "Has the evidence for order A been received?" Compare the follow-up subject against the material subject, and compare the evidence status field in the material. If the material status is "evidence not yet received," the answer should keep that status; if the answer states it has been received or completed, the answer is inconsistent with the material.