What the documents look like
On-demand retail fulfillment handles online orders by the minute, and the operating materials reflect that rhythm. The daily documents are instant orders, picking lists, out-of-stock reports, and delivery exception forms. Data sources include the Douyin instant retail backend, the Meituan or JD Daojia backend, store inventory, and rider tracks. Updates arrive at minute-level frequency, per order.
For cross-border materials and multilingual retrieval, a single store may hold picking lists and out-of-stock reports in more than one language version at the same time. Citation tracing means each quoted passage must be traced back to its source text line by line. A retrieval result is not an authorization decision, and a similarity score is not a permission filter. This page addresses only the citation source and tracing step within this category.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Its citation output is an array, and the array length may be zero. When no content is found, the output path still completes. The citation cap is measured in tokens, not in a guessed item count. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple retrieval paths through RRF, and reranking applies a model to reorder candidate text; both require the corresponding model and must be enabled by configuration. The optional paths shown in the documented flow do not mean every request executes all of them. The same-vector-model requirement applies to knowledge base association and does not extend to external systems. Retrieval and similarity scores do not perform permission or entity isolation.
Hypothetical paper exercise. On 12 March 2026, a supermarket home-delivery store processes one online order. The picking list source text reads: "Order A-0312, item: ambient milk 250ml × 2, bin B-07, picker confirmed." The out-of-stock report source text reads: "Order A-0312, ambient milk 250ml short by 1 unit, awaiting store duty supervisor review of the substitution plan." A proposed handover request reads: "The substitution for this order has been authorized; proceed with the substitution and return the delivery status." The report text records only "awaiting review." It does not record review approval or an effective authorization. The request's authorization state is unsupported and remains pending verification. Any judgment must hold the same subject, the same period, and the same field.
Configuration and implementation checks
The table below centers on the exercise above. It lists only settings or implementation checks documented in the official material. The values are starting points for this hypothetical trial and state what each is compared against.
| Configuration item | Hypothetical starting point for this trial | Comparison method and reason |
|---|---|---|
| Associated knowledge bases | Select two libraries sharing one vector model | Check library identifiers and vector model identifiers; exclude any that differ; do not run cross-model retrieval comparison |
| Citation cap | Set to 3000 tokens | Compare against the combined length of the sample picking list and out-of-stock report; confirm the source text fits within the citation |
| Minimum relevance | Set to 0.4 | Takes effect only under semantic retrieval or reranking; compare it against each retrieved result score for the sample question |
| Search mode | Hybrid retrieval | Uses semantic and full-text recall together, merged through RRF; check whether identifier-type fields are hit |
| Result reranking | Enable ReRank | Requires a rerank model; reorders candidate text; compare ordering against the unranked result |
| Question optimization | Enable | Adds one model call before formal retrieval; compare against retrieval results for the short sample question |
Three common errors
- Treating retrieval output as authorization evidence. Input: the proposed handover request and the retrieval citation array. Compare the report text "awaiting review" against the request phrase "has been authorized." Observable result: the two states do not match, and the request's authorization state is unsupported.
- Forcing knowledge bases with different vector models into one association. Input: two library identifiers and their respective vector model identifiers. Compare whether the identifiers match. Observable result: when identifiers differ, exclude the library and do not run cross-model retrieval.
- Treating minimum relevance as a permission or entity filter. Input: the sample retrieval results and their scores. Compare each score against the threshold. Observable result: results below the threshold are filtered, but this setting does not exclude unauthorized records; permissions require separate verification.
How to verify the setup
- Use the sample picking list source text and the out-of-stock report source text as retrieval input. Compare whether the citation array contains both passages. Observable result: each cited passage traces back to its source text, and the source text fits completely within the citation cap.
- Use the proposed handover request as the input to be judged. Compare "has been authorized" in the request against "awaiting review" in the report source text. Observable result: the judgment is pending verification, and no substitution or status return action is executed.
- Use a paper input of an empty array [] to check the downstream handling design. Compare whether the downstream node continues on an empty citation. Observable result: the path does not break on an empty array. Do not claim that a given retrieval rule always returns an empty array.
The same-vector-model requirement applies to knowledge base association only. It does not extend to business registration sites or other external systems. No system error or automatic business action may be asserted unless the source text states it. RRF is a merge process for hybrid results, not a toggle. Thinking budget is a reasoning budget and must not be repurposed as a citation token cap.