What the documents look like
For product questions in Shopify fulfillment, this page examines FastGPT hybrid retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the hybrid retrieval step
Start the Shopify product test with four details actually written in the proposed order attachment: a summer-limited T-shirt, size L, blue and one piece. These support a question about the purchase, without supplying material or intended-user information. Inspect whether hybrid candidates keep the product name attached to its attributes, then rerank using a complete size-and-colour question. An attribute detached from its product is not usable evidence merely because it appears near the top. Keep purchase details separate from product-care evidence in the review checklist. Finding all four order fields does not supply material or washing instructions. A result can be relevant to the purchased item while still lacking the particular attribute asked about.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | T-shirt name with size L, blue and one piece | Check that attributes still identify the same order item. |
| Complete question | Winter sweater in the order; spring jacket on the label | Preserve both descriptions and file identities. |
| Minimum relevance | Check under semantic-search or reranking conditions | Inspect coverage of both order and label after filtering. |
| Citation token limit | Keep product name, size and colour associated | Trace any omitted size-L or blue wording. |
| Missing-source retest | Order attachment lacks material and washing instructions | Add corresponding product information; avoid another item’s attributes. |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Hybrid retrieval | Vector and full-text results use RRF fusion. |
| Reranking | Reranking reorders candidates using the full question. |
Minimum relevance | Filtering applies to semantic search or reranking. |
Citation limit | Citation content is limited by tokens. |
Question optimization | Compare subjects and restrictions after optimization. |
Three common errors
The example naming a winter sweater in the order and a spring jacket on the shipping label needs two separately identified descriptions. Confirm that both passages were recalled, then inspect whether final citations merge the names into one product. Different text alone does not establish an actual wrong shipment: the order identity and versions still need checking. Reranking operates on existing candidates, so a missing label passage requires a recall-scope check before a ranking-model change.
How to verify the setup
An order attachment containing only name and quantity still lacks evidence for material or washing instructions. Inspect whether the answer borrows attributes from another item, then repeat the test after adding the corresponding product information. Preserve before-and-after candidates and citations for the same question to show when evidence entered the chain. Apply filtering under its documented conditions; a higher score cannot replace the material or care information that has not been supplied.
| Material context | Scope |
|---|---|
| Document types | Order Details, Shipping Rate Template, Shipping Label and Refund Records |
| Candidate sources | Shopify Admin, Payment Gateway, Carrier System and Inventory Location Data |
| Technical baseline | FastGPT v4.17.2 |