What the documents look like
For product questions in Shopify fulfillment, this page examines FastGPT vector indexes. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the vector indexes step
A product-index test for an independent store can use new proposed order wording: “Frosted cup, single-cup package; gift box purchased separately.” Search text for style and packaging must both lead back to this original. The style question can read “frosted”; the packaging question needs the relationship between a single-cup package and a separately purchased box. Two entries identifying the same item do not establish that an optional component is included in the order. Inspect index wording and original meaning separately.
Configuration and implementation checks
| Item or implementation check | Original or index content | Traceable evidence |
|---|---|---|
| Embedding model | Order and specification bases associated with the same retrieval node | Check the same embedding model across bases associated with that node. |
| Original and indexes | Single-cup package; box purchased separately | Do not add an included box in the index. |
| Index content | Style and packaging questions | Inspect the original reached by each. |
| Image path | Capacity only in the missing specification image | Inspect outputs appropriate to the actual text or image path. |
| Change retest | Fixed style, packaging and capacity questions | Compare citations beyond score changes. |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Embedding model | Knowledge bases associated with one search node must use the same embedding model. |
Original data and indexes | One original data block can have multiple indexes. |
Index content | Index wording should be complete and focused. |
Image indexes | Where image retrieval is used, inspect descriptions or image vectors. |
Version and result check | After a change, reuse questions to compare originals. |
Three common errors
The first error says frosted cup with gift box while the original order still says the box is purchased separately. Compare index and order wording and retain that condition without adding inclusion from a packaging query. The second uses a correctly named entry targeting packaging text for another order. Inspect the order identity of the original data, beyond a matching cup name. The third turns separately purchased into a claim that this order did or did not buy the box. Check the actual selected-item record. A condition, an index mapping and a purchase outcome need different evidence; an index hit cannot verify the transaction. Order identity also needs the packaging condition tied to the current transaction, since identical cup names do not establish identical selections. For the purchase outcome, read this order entry rather than reusing the optional-product note as transaction proof. Even a correctly targeted order can describe an option without listing it as selected.
How to verify the setup
If capacity appears only in a specification image that has not been supplied, additional style indexes cannot provide its value. After obtaining the material, determine whether the actual path processes text or images. Compare extracted wording on the text path; inspect descriptions or image vectors only on an image-retrieval path, then trace a capacity question to the corresponding source. Keep the same product questions across model or index changes, recording cited originals and attributes still unsupported. A higher similarity score does not establish coverage of every product parameter. Also inspect an incorrect-target failure: if this order’s packaging question retrieves another order or specification record, a similar name is not evidence for this question. Compare the requested order, index target and returned source identity and version before evaluating the capacity or packaging wording.
| 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 |