What the documents look like
On-demand retail store fulfillment handles online orders on a minute-by-minute cadence, which differs from periodic stocktaking. The daily materials are live orders, picking lists, out-of-stock reports, and delivery exception forms. Sources include Douyin on-demand retail, the Meituan or JD Daojia back office, store inventory, and rider tracking. In cross-border material and multilingual retrieval work, these materials can appear in several languages, and the retrieval targets are usually order numbers, product names, store identifiers, and delivery status fields.
Materials update per order. This page does not derive batch synchronization or transaction-triggered mechanisms from that cadence. Picking lists and out-of-stock reports often carry multilingual descriptions of substitute products, and delivery exception forms may record rider tracking fragments. Model connection and configuration work starts by confirming the language distribution and field structure of these materials, then deciding the capability range of the index model and the language model.
Constraints at the model connections step
Mixed-language material means the index model must cover the languages that actually appear, and the model identifier in a retrieval request must match the vector model already in the library. Per the official documentation, the system needs at least one language model and one index model to work normally, and only one speech recognition model takes effect. AI Proxy is a required dependency. Both the App and Admin services must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail.
The model ID is the value of the model field in the interface request body and is globally unique. Model mapping can map the model FastGPT requests to a specific upstream model, for example mapping gpt-4o-test to gpt-4o. A retrieval relevance score does not carry permission filtering or entity isolation duties, and retrieval itself is not authorization. Libraries built on different vector models cannot be associated together; check the model identifier first and exclude those that do not match.
Hypothetical drill sample: a store fulfillment team receives a cross-border picking document with three records. Record one: "Order number ORD-2026-0417, store SH-03, product 500ml mineral water, picking status: picked." Record two: "Order number ORD-2026-0417, store SH-03, product 500ml mineral water, delivery status is awaiting pickup." Record three: "Order number ORD-2026-0418, store SH-03, product 1L milk, out-of-stock report: pending review." The proposed handover request reads: "Please confirm that delivery for ORD-2026-0417 is complete, and authorize out-of-stock handling for ORD-2026-0418."
This request involves two different orders. Record two shows only awaiting pickup, and record three's out-of-stock report is pending review. Pending review is not approval, and awaiting pickup is not completed delivery. The authorization portion of the handover request has no support in the sample.
Configuration and implementation checks
The configuration below centers on this page's multilingual retrieval sample and lists settings or implementation checks that the official documentation describes. The values are starting points for this hypothetical trial, not official defaults.
| Configuration item | Starting point for this hypothetical trial | Reason for the choice and comparison method |
|---|---|---|
| Index model | Select one index model that supports multilingual text | Compare against the languages present in the sample, and compare its identifier with the vector model already in the library to confirm they match |
| Language model | Select one language model that supports text dialogue | Check the input types the application needs, and compare them with the model capability fields |
AIPROXY_API_ENDPOINT | Enter the AI Proxy service root address | Per the documentation, do not append /v1; compare with the actual deployment address |
AIPROXY_API_TOKEN | Enter a non-empty token consistent with the server side | Compare with the AI Proxy administrator token and confirm they match |
| Model mapping | Map the requested model identifier to the actual upstream model | Compare against the model name in the channel configuration and confirm the mapping relationship |
| Retrieval scope | Limit retrieval to the library holding the sample material | Check the library's vector model identifier and exclude libraries on a different model |
Three common errors
- Treating a retrieval hit as authorization evidence. In the sample, the out-of-stock report for ORD-2026-0418 is pending review, and retrieving that record does not mean out-of-stock handling has been approved. The input is the retrieval result, the comparison object is the report status field, and the observed result is that the status remains pending review.
- Merging the status of different orders into one order. In the sample, ORD-2026-0417 and ORD-2026-0418 are different orders; record two's awaiting pickup belongs to the former, and record three's pending review belongs to the latter. The input is the two records, the comparison object is the order number field, and the observed result is that the order numbers differ.
- Treating awaiting pickup as completed delivery. Record two shows only awaiting pickup and states no delivery time or signature field. The input is record two, the comparison object is the delivery status field, and the observed result is that no completion marker appears.
How to verify the setup
- Check the index model against the library's vector model identifier. The input is the model identifier of the library holding the sample material, the comparison object is the identifier of the index model in the configuration, and the observed result is that the two match. If they do not match, record it as pending verification and do not run a cross-model retrieval comparison.
- Check whether
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare both set. The input is the environment variable configuration, the comparison object is the documentation's required conditions, and the observed result is that both items are non-empty. If either is missing, record that the startup condition is not met. - Check whether the authorization portion of the proposed handover request has corresponding evidence. The input is the draft request and the three sample records, the comparison object is the two items "delivery is complete" and "authorize out-of-stock handling," and the observed result is that record two shows no completion and record three is pending review. Neither item is supported, so the evidence gap stays open and no authorization clause is supplied.