What the documents look like
On a Shopify standalone store, the merchant configures payment, shipping, and fulfillment services, and the platform does not run one unified automatic return process. Order details, shipping templates, shipping labels, and refund records therefore sit across Shopify Admin, payment gateways, carrier systems, and inventory location data. Updates follow a per-order and daily rhythm, so a retrieved fulfillment status can lag behind the current state. Independent-store merchants, Shopify operators, and third-party fulfillment providers share one knowledge base, and retrieval quality across multilingual and cross-border material shapes how payment disputes, logistics complaints, and overselling are handled.
A knowledge base search node can link one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together; the identifier is checked first and non-matching bases are excluded. The node outputs cited content as an array whose length can be zero, and the downstream path still runs when nothing is found. Semantic search relies on vector distance and supports cross-language understanding, which suits natural-language questions. Full-text search relies on keyword hits and suits numbers, model codes, proper nouns, and error codes. Hybrid search uses both recall paths and merges them through RRF. The minimum relevance value is a filter threshold that applies only in semantic search or when result reranking is used, and it does not act as a permission or entity isolation mechanism. The citation limit controls, in tokens, the maximum content cited per search. Question optimization performs coreference resolution and question expansion before retrieval, adding one model call and overall latency. Image retrieval requires the corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. The optional paths shown in the overall flow diagram do not mean every request executes all of them.
Constraints at the retrieval step
Hypothetical drill sample: an order detail record in a standalone merchant knowledge base reads "Order A-1042, placed 6 March 2025, delivery address Germany, payment status paid, fulfillment status awaiting shipment, carrier not assigned." A shipping template record reads "Germany standard shipping template, applies to orders below 150 EUR, transit 5 to 8 business days." A draft sentence awaiting review reads "Order A-1042 has been collected by the carrier, and the refund request has been approved." The review compares this draft sentence against the two source records on fulfillment status and refund status.
Because the search node can link only knowledge bases sharing one vector model, a cross-model link is not performed and the identifier check comes first. Because the output array can be empty, an empty citation list is a valid downstream input rather than a retrieval failure. Because the minimum relevance value applies only in semantic search or with reranking, it cannot be relied on to exclude unauthorized orders. Because question optimization adds a model call, a multi-turn follow-up that contains a reference is compared against its optimized form to check whether the subject was altered.
Configuration and implementation checks
| Configuration item | Hypothetical drill starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the order and shipping material bases that share one vector model | Check the vector model identifier first, exclude non-matching bases, and do not run cross-model retrieval |
| Search mode | Run the same question once in semantic search and once in full-text search | Compare which content each mode returns for an order-number style question |
| Citation limit | Set a token amount that can hold one full order record and one shipping template | Test with material of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Set the threshold after labelling relevant and irrelevant samples | Applies only in semantic search or reranking; observe whether unrelated fragments are filtered |
| Question optimization | Enable for multi-turn follow-ups that contain references | Compare the original question with the optimized question and check whether the subject was changed |
| Result reranking | Enable when candidates are numerous and a rerank model is available | Compare ordering before and after reranking and observe whether the score is used for filtering |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Input a query for order A-1042 and compare whether returned content includes other orders or unauthorized records. A relevance score reflects textual similarity only and does not perform business isolation; unauthorized content must be blocked by a downstream permission implementation.
- Linking knowledge bases that use different vector models. Input two bases with different vector model identifiers and compare whether the node allows both to be selected. The documentation states that only bases with the same vector model can be linked; when identifiers do not match, exclude them and do not go on to compare missed recall or redundant results.
- Treating "collected by the carrier" and "refund request approved" in the draft sentence as established facts. Input the draft sentence and the order source record, and compare the fulfillment status field. The source states awaiting shipment and carrier not assigned, so the collection description in the draft is unsupported; no refund record is provided, so the refund status remains to be verified.
How to verify the setup
- Input the order source record and the draft sentence from the hypothetical sample, and compare the fulfillment status field. The source reads awaiting shipment and carrier not assigned, while the draft says collected; these are mutually exclusive values in the same field, so the draft sentence is judged unsupported. The expected observation is that the sentence is marked in the review record as inconsistent with the source.
- Input the order number query for A-1042 and run it once in semantic search and once in full-text search, comparing whether the returned content includes that order record. The expected observation is that each mode returns candidates, with the actual hit situation left to observation; a first-position hit or full recall is not guaranteed.
- Input an explicitly empty citation array [] and check whether the downstream flow continues to run. The documentation states that cited content can have length zero and the output path still runs; the expected observation is that the downstream node receives the empty array and continues as designed, without claiming that a fictional rule search necessarily returns an empty result. All of the above is a paper review; no deployment or test has taken place.