What the documents look like
A Shopify standalone store leaves payment configuration, shipping rates and fulfillment service to the merchant, and the platform does not run a single automatic return process. That difference shapes after-sales work: the materials that answer a buyer question are assembled by the merchant from several systems rather than pulled from one marketplace record.
Four document types carry most of the load. Order details come from Shopify Admin. Shipping templates describe the rate structure the merchant configured. Shipping labels come from the carrier system. Refund records sit alongside payment gateway data and inventory location data. A single after-sales question can cross all four, for example when someone asks what state a particular order is in or which shipping label a refund record belongs to. These materials update per order and daily, so a retrieval step may be answering against records that changed since the last review.
Retrieval and recall do not change order state. They pass already-indexed reference content to a downstream model. A retrieval hit is not business authorization, and it does not mean a refund, reshipment or termination has occurred.
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. Check the vector model identifier first and exclude anything that does not match; do not run a cross-model search to compare missed recalls. The smallest search unit is the knowledge base. Collections exist for classification management and do not affect search results.
A search is not a direct path from question to vector store. The documented options include question optimization, semantic search, full-text search, hybrid search, image search, RRF fusion, ReRank secondary ordering, similarity filtering and a reference token cap. These combine according to input content and search parameters. Image search requires image input and matching model capability. Reordering requires a model and must be enabled in settings. The flow diagram lists available paths; it does not mean every request executes all of them. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reordering is used. Reference content is output as an array whose length can be zero, so the output path completes even when nothing is found.
Hypothetical paper test case
The following is a hypothetical paper test case for checking purposes. It does not come from any real customer or real record.
Source text one: "Order number H-2041, placed 4 March 2025, one coat, delivery address in a named province and city."
Source text two: "The shipping label for this order was generated on 6 March 2025, and the carrier pickup record shows collection on 7 March 2025."
Source text three: "Refund record R-118 is linked to order H-2041, registered 10 March 2025, current status pending review."
Draft sentence to be assessed: "The refund for order H-2041 is complete and the buyer can be told the funds have arrived."
Configuration and implementation checks
The table below covers settings the official documents describe, or implementation checks tied to this page's sample. The values are starting points for this hypothetical trial, not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Order details base and refund records base | Both use the same vector model; check the vector model identifier and exclude any mismatch |
| Search mode | Hybrid search | The sample contains identifier H-2041 and natural-language wording; compare against semantic-only and full-text-only results |
| Reference cap | 1200 tokens for this hypothetical trial | The sample holds three short records; compare against 600 and 2400 to see whether reference content is truncated |
| Minimum relevance | 0.35 for this hypothetical trial | Label relevant and irrelevant samples first, then adjust the threshold; it applies only under semantic search or result reordering |
| Question optimization | Enabled | The sample contains references and an identifier; compare the original question with the optimized question and check whether the subject was altered |
| Result reordering | Enabled per settings | Secondary ordering when candidates are numerous; compare ordering against results without reordering |
Three common errors
- Treating a retrieval hit as authorized execution. The input is the draft sentence and the retrieved refund record R-118. Compare the record's status field, pending review, with the draft sentence's claim that the refund is complete. The observable result is that the two do not agree. Pending review does not show that a refund was executed, and it does not show that an operator lacks permission.
- Associating knowledge bases across vector models. The input is the vector model identifier of the order details base and of the refund records base. Compare whether the two identifiers match. The observable result is that only matching identifiers can be associated together. When they differ, record the actual observation; do not infer missed recalls or redundant returns.
- Using a similarity threshold in place of entity separation. The input is the same question run against records for unrelated orders. Compare the minimum relevance setting with the returned results. The observable result is that relevance filtering acts on scores only; it does not exclude unrelated business or unauthorized records. Retrieval itself carries no permission decision.
How to verify the setup
- Against source texts one and two, enter the question "Has H-2041 been collected?" Compare whether the returned references contain both the shipping label generation date and the carrier collection record. The observable result is that both records appear in the reference content without truncation. This is the expectation; it has not been tested.
- Against source text three and the draft sentence, enter the draft sentence. Compare the draft claim that the refund is complete with the record status pending review. The observable result is a finding that the claim is unsupported, and the evidence gap stays open. Do not supply authorization or effective clauses to close it.
- Against empty-result handling, enter the paper input []. Compare whether the reference array length is zero. The observable result is that the output path completes and the downstream step can accept an empty reference. Do not claim that a fabricated rule search necessarily returns [].