What the documents look like
A Shopify storefront is configured by the merchant, so payment, shipping and fulfillment arrangements are not standardized by a platform-level returns process. Fulfillment materials therefore arrive from several sources: order details in Shopify Admin, settlement records from the payment gateway, shipping labels from the carrier system, inventory location data, and refund records. These materials are commonly linked by identifiers such as order number, waybill number and refund number. Identifiers suit full-text retrieval, while clauses describing responsibility and natural-language passages suit semantic retrieval.
The per-order and daily update rhythm means materials for one order can exist in more than one version at different stages. The comparison task is to judge whether two materials agree on subject, object, status or effective scope. The retrieval step only recalls relevant materials; it does not decide which version is valid.
Constraints at the retrieval step
The smallest search unit is the library. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together. When this situation appears, check the vector model identifier first, exclude the non-matching library, and record what was observed. Do not run a cross-model search to compare missed recalls.
One retrieval combines question optimization, semantic retrieval, full-text retrieval or hybrid retrieval, image retrieval, RRF fusion, and ReRank, followed by similarity filtering and a citation cap. These paths are not all executed on every request. Image retrieval requires a corresponding image input and model capability. ReRank requires a model and must be enabled in the settings. The citation cap is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not provide permission or entity isolation. Retrieval itself is not authorization. Forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation.
Hypothetical paper exercise sample: a storefront merchant checks refund responsibility for one order with a third-party fulfillment provider. The source text supplied for comparison is three sentences. "The refund record for order A states a refund amount of 200 yuan, a refund initiation date of 10 March 2025, and a status of pending review." "The payment gateway settlement record for the same order A states a settlement amount of 200 yuan for this transaction, with a settlement date of 12 March 2025." "Clause 4 of the service agreement between the parties states that refund responsibility is divided according to the fulfillment stage of the order at the time the refund was initiated." The draft sentence to be judged is: "The 200 yuan refund for order A has been borne by the fulfillment provider, and the refund is complete." This draft sentence concerns the bearing party and the completion status. The source text does not state a bearing party, and the refund record status is pending review.
Configuration and implementation checks
| Configuration item | Purpose on this page | Value and basis |
|---|---|---|
| Associated knowledge bases | Limit the retrieval scope for comparison materials | Select knowledge bases containing order details, refund records and the service agreement; each library must use the same vector model, checked one by one against the identifier |
| Search mode | Choose a recall path by question type | Identifier questions use full-text or hybrid retrieval; natural-language responsibility questions use semantic or hybrid retrieval |
| Citation cap | Control the length of cited content in one retrieval | Hypothetical starting value for this exercise is 2000 tokens, chosen because the three source sentences in this sample are short; compare against 1000 and 4000 to see whether cited content stays complete |
| Minimum relevance | Filter low-relevance results | Hypothetical starting value for this exercise is 0.3, effective only in semantic retrieval or result reranking; label relevant and irrelevant samples before adjusting |
| Question optimization | Handle references and short questions | The sample question contains "this order" as a reference; after enabling, check whether the optimized result keeps the order A subject |
| Result reranking | Secondary ordering of candidate text | Requires a ReRank model and must be enabled in the settings; the rerank score is 0-1 and can be used for filtering |
Three common errors
- Treating the refund record status "pending review" as a completed refund. This field is a status value; pending review does not mean the refund has been executed, and the draft sentence's "the refund is complete" is not supported by the source text. Check: input the refund record source text and the draft sentence, compare the status field value, and observe whether the draft sentence agrees with "pending review".
- Treating a retrieval hit as a basis for identifying the bearing party. Retrieval only recalls materials; responsibility requires evidence linking the agreement clause to the fulfillment stage. Check: input clause 4 of the service agreement and the draft sentence, compare whether the agreement states a bearing party for this order, and observe whether source text supports "borne by the fulfillment provider".
- Comparing dates from different fields directly. The refund initiation date and the settlement date belong to different fields; a difference between them does not constitute a contradiction. Check: input both date records, compare the field names and values, and observe whether they are mutually exclusive values of the same field.
How to verify the setup
- Input the three source sentences and the draft sentence from the sample, compare the bearing party and completion status in the draft sentence against what the source text states. The draft sentence is supported only when the source text clearly states a bearing party and the refund status is complete. The current sample should be judged unsupported, with the bearing party still to be verified.
- Input "order A refund" as the retrieval term and observe whether the returned content includes the refund record source text. If the return is empty, use the explicit paper input
[]to check whether the downstream design continues, without claiming that this query necessarily returns empty. - Input a follow-up question containing "this order", compare the retrieval terms before and after question optimization, and observe whether the optimized result keeps the order A subject. If the subject is changed incorrectly or lost, record the observation and adjust the question optimization setting.
This page addresses only Shopify storefront order fulfillment / contract and transaction material comparison / knowledge base retrieval and recall. During operational handover, the authorization status of materials, operations and requests must be established separately; retrieval or a form does not constitute authorized execution.