What the documents look like
For product questions in On-demand retail fulfillment, this page examines FastGPT vector indexes. A substitute item does not automatically inherit the original order ingredient description. The examples are hypothetical.
Constraints at the vector indexes step
A proposed instant-retail packet orders an original-flavour biscuit, suggests asking about a wholegrain alternative and refers the original recipe to its own text. A question about the current item being wholegrain lacks acceptance or actual picking of the alternative. Original-flavour recipe material cannot supply the alternative ingredients. An entry merging both into original wholegrain biscuit would create an unsupported name. Preserve order, suggestion and recipe subjects, without treating related food categories or discussion within one order as proof that the purchased item changed.
Configuration and implementation checks
| Original-data/index check | Specific subject relationship | Acceptance action |
|---|---|---|
| Original order identity | Original-flavour biscuit | Do not change to purchased wholegrain. |
| Alternative state | May ask | Seek actual selection evidence. |
| Recipe subject | Each item has its own text | Do not borrow another recipe. |
| Mapping separation | Do not merge into a new product name | Inspect originals individually. |
| Change boundary | New data does not establish order change | Check citations and actual event. |
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
First, a substitute suggestion may be indexed as purchased wholegrain; inspect selection or order-change records. Second, a wholegrain question may target the original recipe; inspect product identity and applicability. Third, after alternative-product data is added, entries may still combine both items; inspect each original and the names in actual citations. These require business choice, recipe subject and mapping-separation checks. Requests and similar names cannot supply nutrition, taste or actual delivery conclusions, and an available alternative cannot become an accepted replacement simply because it appears near the order text.
How to verify the setup
Acceptance should ask what the order names, which alternative is suggested and whether a change is evidenced. Keep citation subjects separate and retain a possibility without claiming buyer acceptance. Ingredient questions need the requested item original rather than another recipe. After adding data, inspect actual processing and indexes before same-question target comparison. Upload success does not establish an order change. Several entries for one original still establish that original content only; a substitution relationship cannot make the two products properties equivalent or establish what was physically supplied.
| Material context | Scope |
|---|---|
| Document types | Instant Order, Picking List, Out-of-Stock Report and Delivery Exception Order |
| Candidate sources | Douyin Instant Retail, Meituan or JD Daojia Backend, Store Inventory and Rider Tracking |
| Technical baseline | FastGPT v4.17.2 |