What the documents look like
For product questions in On-demand retail 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
The proposed instant-order attachment lists fresh milk, whole-fat specification and two units: name, variant and quantity. Candidates for a basic-product question should keep all three attached to the same item before reranking. The whole-fat label does not supply nutritional figures or an intended user group, and “fresh” in a name does not verify actual quality. The answer needs to trace its description to the order wording within the scope of a document record.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | Fresh milk, whole-fat variant and two units | Keep name, variant and quantity tied to one item. |
| Complete question | One ordered box versus two picked boxes of eggs | Check units and linkage without asserting an extra picked box. |
| Minimum relevance | Inspect filtering under semantic or reranking conditions | A candidate quantity does not establish delivered quantity. |
| Citation token limit | Both original quantities and source roles | Do not silently change one value to the other. |
| Missing-source retest | Order lacks ingredients and shelf-life evidence | Do not attribute another product’s information to this order. |
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
One box of organic eggs in the order and two in the picking list require unit, product and order-linkage checks. Confirm both quantity passages in candidates, then inspect reranking and citations for either omission. Versions, split handling or other records may affect interpretation, but this example provides no such evidence. Preserve the difference for review without declaring an extra box was picked or selecting an actual completed quantity by rank. Trace the ordered requirement separately from the picking record, without silently changing either quantity to the other. A later question about the amount actually delivered requires corresponding delivery evidence. Selecting one of these two document values cannot establish that later event.
How to verify the setup
The existing order contains neither an ingredient list nor shelf-life information. Name and quantity cannot replace these sources. Test related questions for product information retrieved from another item and incorrectly attributed to this order. After obtaining corresponding material, inspect the added passage through filtering, ordering and citations while preserving its conditions. High relevance does not establish complete attribute verification; examine which statements the actual evidence supports one by one.
| 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 |