What the documents look like
On a Shopify standalone store, the merchant configures payment, shipping and fulfillment services directly; the platform does not run a single automatic returns process across all stores. The documents that matter for copy drafting are therefore merchant-held records rather than platform-issued ones: order details, shipping rate templates, shipping labels and refund records. These update per order and daily. A product or marketing draft that states a delivery window, a free-shipping threshold or a return condition is only supported to the extent the matching document states it. Order details carry the item name and order scope; a shipping rate template carries the regions and thresholds it covers; a shipping label carries the carrier handover date; a refund record carries a processing status. The drafting task in this page is to decide whether a proposed copy sentence is supported by those documents, and to keep unsupported statements marked as unsupported rather than filled in.
Constraints at the multi-turn prompts step
The chat interface takes chatId. When chatId is a non-empty string, the conversation is read from the database and only the last element of the messages array is used as the user question; the remaining messages are ignored. When chatId is absent, no stored context is used and the context is built entirely from the supplied messages. Memory turns control how many earlier rounds the AI may reference; a larger value brings in more prior content. Reply length caps a single answer. A lower temperature suits rule-bound tasks; a higher temperature suits creative writing. Under reply format, choosing json_schema requires the corresponding format requirement to be filled in. Multimodal recognition depends on the model's own capability, and ordinary document files are not sent directly to the model as multimodal input; document content still has to be parsed into text. Thinking configuration aligns with reasoning_effort, and the thinking budget is a control of thinking intensity, not a quotation-token limit.
For copy drafting, if the output must follow a fixed structure, json_schema is the relevant choice and its format requirement must be supplied. If the draft needs to reference selling points confirmed in earlier rounds, the memory-turn count has to be checked against the rounds actually required. If a draft describes image material, the model must support image recognition and that capability must be enabled. A shared embedding-model requirement applies only to knowledge-base association and does not extend to external systems.
Configuration and implementation checks
Hypothetical paper exercise sample. Source text: "This store's shipping rate template states: East China region, free shipping on orders of 199 yuan or more, otherwise a 12 yuan shipping fee. The shipping label shows order A was handed to the carrier on 10 March 2025. The refund record shows order B has status pending review." Draft sentence to be judged: "All orders in this store ship free over 199 yuan and are dispatched within 24 hours of payment." The task is to decide whether the draft sentence is supported by the source text.
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Memory turns | Set to 4 rounds | The sample covers shipping rate, dispatch and refund material across earlier rounds; compare with 2 rounds to observe whether shipping rate detail is lost |
| Reply length cap | Set to 800 | The draft sentence is short and 800 can hold a full support judgment with evidence references; compare with 200 to observe truncation of evidence |
| Temperature | Set to 0.3 | The judgment task requires stability; compare with 0.8 to observe whether statements beyond the source text appear |
| Reply format | Select json_schema | The judgment needs fixed fields separating supported and unsupported; compare with the default format to observe whether downstream can parse it |
| Multimodal recognition | Image recognition off | The sample material is text; compare with image recognition on to observe whether unrelated image parsing is introduced |
| Thinking configuration | Select standard thinking | The judgment compares three documents; compare with no thinking to observe whether the refund status is missed |
Three common errors
- Treating "all orders in this store ship free over 199 yuan" as supported. The source text states only the East China region at the 199 yuan threshold and states nothing about other regions. Input is the source text and the draft sentence; the comparison is between "East China region" and "all orders"; the observable result is that the draft sentence widens the scope, so it is unsupported.
- Treating "dispatched within 24 hours of payment" as supported. The source text states only that order A was handed to the carrier on 10 March 2025 and states no dispatch time limit after payment. Input is the shipping label and the draft sentence; the comparison is between the carrier handover date and "within 24 hours of payment"; the observable result is that the source text does not state that time limit, so it is unsupported.
- Treating "pending review" in the refund record as a completed refund. Pending review is an in-progress status and the source text does not state the review outcome. Input is the refund record and the draft sentence; the comparison is between "pending review" and "completed"; the observable result is that the statuses differ, so it is unsupported.
How to verify the setup
- Input the three source sentences and the draft sentence, and compare "all orders in this store ship free over 199 yuan" against "East China region, free shipping on orders of 199 yuan or more". Observe whether the output marks the scope difference as unsupported. If the output marks it as supported, the configuration does not meet the judgment requirement.
- Input the three source sentences and the draft sentence, and compare "dispatched within 24 hours of payment" against "handed to the carrier on 10 March 2025". Observe whether the output marks the missing time limit as unsupported. If the output supplies a time limit, the configuration does not meet the judgment requirement.
- Input the three source sentences and the draft sentence, and compare the refund wording in the draft against "pending review" in the refund record. Observe whether the output marks pending review as not completed. If the output marks it as completed, the configuration does not meet the judgment requirement.
For an empty-result handoff, a clearly labelled paper input of [] can be used to check the downstream design; no claim is made that a fictional rule search will return []. The retrieval check observes the failure branch and does not guarantee a first hit or full recall.