What the documents look like
A Shopify standalone store sets its own payment, shipping and fulfillment services, and the platform does not run one automatic return process across all merchants. Order details, shipping-rate templates, shipping labels and refund records therefore sit in separate places: Shopify Admin, payment gateways, carrier systems and inventory-location data. An operating review or an exception review has to reconcile the same order across those sources.
The update rhythm given for this category is per order and daily. Retrieval at this step works only on material that has already entered a knowledge base. Whether the material has finished training, and whether the selected knowledge bases share one vector model, determine whether a search can run at all.
Constraints at the retrieval step
The knowledge-base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together in one search. The vector-model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. Search output is an array of quoted content, and its length can be 0. When nothing is found, the output path still runs. Retrieval is a recall action, and it does not confirm the status of an order, a refund or an authorization.
Hypothetical drill sample. Material text available for comparison: the order-detail record states "Order A was created on 10 March 2025, payment status paid, fulfillment status unfulfilled"; the refund record states "Order A submitted a refund request on 12 March 2025, status pending review"; the shipping-rate template record states "The East China region shipping-rate template has been effective since 1 March 2025." Draft sentence to be judged: "Order A has completed the refund, and the fulfillment exception is closed." That draft sentence does not agree with the status in the refund record, which is pending review. Pending review does not amount to a completed refund.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only one order-fulfillment base that uses the same vector model | Check the vector-model identifier first, exclude non-matching bases, then run the search |
| Search mode | Full-text search for identifier-type questions; hybrid search for natural-language review questions | Compare semantic, full-text and hybrid search against the test question; no single mode is assumed to suit all material |
| Quote cap | Hypothetical starting point of 2000 tokens | Test with order details and refund records of different lengths to see whether content stays complete, then adjust the value |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first; this value takes effect only in semantic search or when result reranking is used |
| Question optimization | Enabled for multi-turn follow-ups that contain references | Compare the original question with the optimized question and check whether the subject was changed incorrectly |
| Result reranking | Enabled per settings when candidate results are numerous | Reranking requires a model and must be enabled in settings; its score is 0-1 and filtering uses that score |
Three common errors
- Treating "pending review" in the refund record as a completed refund. The input is the refund record text above; the comparison is between "has completed the refund" in the draft sentence and the status field in the source text; the two values differ, so the draft sentence is not supported.
- Linking knowledge bases that use different vector models. The input is the vector-model identifier of two knowledge bases; the comparison is whether the identifiers match; when they differ, the bases cannot be linked together, and no claim is made here about missed recall or redundant results.
- Treating a retrieval hit as an authorization or a completed business action. The input is the quoted content returned by the search; the comparison is whether it contains an approval or execution record; quoted content shows only that material exists, while refunds, synchronization and similar business actions require downstream implementation.
How to verify the setup
- Input the order-detail, refund-record and shipping-rate template text above, and compare the refund status field in the draft sentence with the status field in the refund record. The condition for "has completed the refund" is met only when both places carry the same mutually exclusive value and point to the same order and the same period. The current source text says pending review, so the judgment is not supported.
- Input the vector-model identifiers of two knowledge bases and compare whether they are the same. The condition for linking them together is met only when they match. When they differ, exclude that base and record the actual observation; do not run a cross-model search in order to compare missed recall.
- Input a paper query with no corresponding material and an empty result array [], then compare whether the downstream node still runs the output path when the quoted length is 0. The condition for handling an empty reference is met when the path still runs. This check does not show that any invented rule search will necessarily return [].
Pending review does not show a lack of operating permission. Signing, reaching cooperation and a contract taking effect are three separate facts. Where evidence is missing, the gap is kept, and anything not stated, or supported only by a related stage, is recorded as pending verification.