What the documents look like
On a Shopify standalone store, the merchant configures payment, shipping and fulfillment services, and the platform does not run a unified automatic return process. This shapes the material that product and marketing copy drafting depends on. The relevant records sit in separate sources: order details in Shopify Admin, payment gateway records, shipping labels from carrier systems, inventory location data, refund records and shipping rate templates. Copy drafting retrieves quotable wording from these materials, for example the actual shipping scope of a product, the fulfillment state of one order, or the handling conclusion attached to one refund record. Data updates per order and daily, so retrieved content can change as an order progresses. A draft sentence that cites stale or inapplicable fulfillment information can contribute to payment disputes or logistics complaints.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Retrieved references are output as an array, and the length can be 0, meaning the downstream chain still runs when nothing is found. Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF, then applies ReRank for a second ordering. The minimum relevance value takes effect only in semantic search or when result reranking is used. The reference limit is measured in tokens and controls how much content each search may cite. Question optimization performs coreference resolution and question expansion, adding one model call before retrieval.
For product and marketing copy drafting, these constraints mean the drafter retrieves the source wording from order details, shipping rate templates, shipping labels and refund records that relates directly to the copy claim, then checks each proposed sentence against that wording. Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to any external system. Retrieval is the process of obtaining reference content, and it does not constitute authorization or business confirmation.
Hypothetical paper test case: a standalone store has one order. The order details record "Order A-1001, product is a cotton short-sleeve T-shirt, order date 1 March 2025, shipping address in Oregon, United States." The shipping rate template records "US domestic orders over USD 50 ship free; this order's product amount is USD 45." The shipping label records "Handed to carrier on 3 March 2025, tracking number TRK-2002." The refund record is empty. One proposed sentence to be judged is: "This product ships free anywhere in the United States." The check compares the free-shipping condition in the rate template, which requires USD 50, against the order amount of USD 45. The material does not support the wording "ships free anywhere."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Order details base, shipping rate template base, shipping label base, refund record base | The four bases use the same vector model and cover the material types this task needs |
| Search mode | Hybrid search | Copy checking needs both semantic matching and exact hits on order numbers and amounts |
| Reference limit | Hypothetical starting point of 3000 tokens for this trial | Compare against actual materials of different lengths and observe whether cited content is complete |
| Minimum relevance | Hypothetical starting point of 0.3 for this trial | Label related and unrelated samples first, then compare filtering results |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare the original question with the optimized question to see whether the subject was changed incorrectly |
| Result reranking | Enabled for this trial | Requires a ReRank model; compare ordering changes after enabling it in the settings |
Three common errors
- Turning a retrieved order detail sentence "this order has shipped" into "all orders have shipped." The input is the order detail text, the comparison is between the order-number scope in the source and the subject scope of the draft sentence, and the observable result is whether the draft sentence exceeds the single order the source limits it to.
- Dropping the condition in the shipping rate template "orders over USD 50 ship free" and drafting "this product ships free." The input is the rate template text, the comparison is between the free-shipping condition and whether the draft sentence includes that condition, and the observable result is whether the draft sentence omits the amount threshold.
- Reading an empty refund record as "this order has no refund risk." The input is the refund record retrieval result, the comparison is between the empty result and the assertion scope of the draft sentence, and the observable result is whether the draft sentence treats missing evidence as a negative conclusion.
How to verify the setup
- Use the order detail text from the hypothetical sample, "Order A-1001, product is a cotton short-sleeve T-shirt, order date 1 March 2025, shipping address in Oregon, United States," as the retrieval input. Compare whether the returned reference contains that order number and product name. The condition is met when the cited content matches the source and is not truncated.
- Use the shipping rate template text from the hypothetical sample, "US domestic orders over USD 50 ship free; this order's product amount is USD 45," as the retrieval input. Compare whether the returned reference contains the amount condition. The condition is met when the cited content presents the free-shipping threshold completely and the draft sentence "This product ships free anywhere in the United States" is judged unsupported by the material.
- Use the empty refund record from the hypothetical sample as the retrieval input. Compare whether the returned reference array has length 0. The condition is met when the downstream design can carry an empty result and does not interpret it as "no refund risk" or "confirmed no refund."