What the documents look like
On-demand retail store fulfillment processes online orders by the minute, which separates it from periodic stocktaking. Orders arrive from Douyin on-demand retail, the Meituan or JD Daojia back office, and store inventory together with rider tracking forms the fulfillment basis. Records update at minute level, once per order. Everyday material includes the immediate order, the picking list, the out-of-stock report, and the delivery exception note. After-sales answers need each statement in the reply traced to the original text of these materials, so source citation becomes the starting point for checking.
A store-to-home supermarket, a rider dispatcher, and a picker all work from the same order record at different moments. When a reply about a substitution or a delivery status does not match the underlying document, the consequence can be a delivery timeout payout, a false-delivery penalty, or a customer refund. That is why the citation step matters more here than the fluency of the answer.
Constraints at the citation tracing step
The knowledge base search node first checks which knowledge bases are selected and which material version they hold. Only knowledge bases that share the same vector model can be associated together. The search node outputs citations as an array, and the length may be zero, which means the downstream path still runs even when nothing is retrieved. Citation content then passes to the answering step, where each answer statement is compared against the cited original text. Traceable samples are used for the display check; a thinking process or a smooth answer does not replace source evidence. Material timeliness is checked by testing sources against the current version and a historical version, and whether a citation is adopted still depends on review of the effective version. Retrieval and relevance scores do not carry permission or entity isolation, and a retrieval result does not constitute authorization.
Hypothetical drill sample: an immediate order A-0630 at a store-to-home supermarket on 3 June 2025 at 14:20. The picking list records "bagged rice 5kg out of stock, reported for substitution with same-specification 5kg pack." The out-of-stock report on the same day at 14:25 records "substitution pending store duty confirmation." The delivery exception note on the same day at 14:40 records "rider has collected goods, in delivery." A draft sentence for judgment reads: "The out-of-stock substitution for this order has been confirmed by the store, and the customer can be told the substitution is complete." The materials show the substitution is still pending confirmation, so the draft sentence turns a pending state into a completed one, and the source does not support that conclusion.
Configuration and implementation checks
| Configuration or check item | Purpose on this page | Correspondence with the sample |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases used for this retrieval | Check which library and version hold the order, picking, report, and exception note |
| Search mode | Semantic search, full-text search, or hybrid search | Order number and product name go through full-text or hybrid search |
| Citation limit | Maximum tokens cited per search | Hypothetical trial starting value set to 2000 tokens because the sample material is short; compare with the displayed value and adjust |
| Minimum relevance | Filter low-relevance results | Takes effect only under semantic search or result reranking |
| Result reranking | Second ordering of candidate text | Requires a rerank model and must be enabled by setting |
| Question optimization | Coreference removal and question expansion | Enabled for multi-turn follow-ups, adding one model call |
RRF is the fusion process for hybrid search results and is not a switch configuration. Thinking budget is a reasoning budget and has no relation to the citation token limit.
Three common errors
- Treating a retrieval output as authorization. In the sample, the out-of-stock report is in a pending-confirmation state. A retrieval hit on that document does not mean the store has confirmed the substitution, and the confirmation record must be checked separately before any reply.
- Treating an empty citation array as a workflow failure. The search node output length may be zero. A paper input of [] can be used to check the downstream handling design, without claiming that a particular retrieval necessarily returns an empty array.
- Associating knowledge bases that use different vector models. Cross-model association is not allowed. Check the identifier first and exclude non-matching bases, rather than running a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the original text of immediate order A-0630, the picking list, the out-of-stock report, and the delivery exception note from the sample. Compare the draft sentence "the out-of-stock substitution has been confirmed by the store" against the report line "substitution pending store duty confirmation." The report does not state a confirmation conclusion, so the draft sentence is unsupported and the evidence gap must be recorded.
- Input the original text of the out-of-stock report from the sample. Compare "substitution pending store duty confirmation" with "confirmed" for the same subject, the same period, and the same field. The two are mutually exclusive values of the same field; the sample takes the pending value, and writing it as confirmed in the draft sentence forms a contradiction.
- Input the paper empty array []. Compare whether the downstream answering step still runs when the search node output length is zero. Observe whether the citation handling path continues as designed, without claiming that a fictional retrieval necessarily returns an empty array.
This page addresses only on-demand retail store fulfillment, order fulfillment and after-sales answering, and citation source tracing. At operational handover, the authorization state of materials, operations, and requests must be checked, and retrieval or a form must not be treated as authorized execution.