What the documents look like
On-demand retail fulfillment processes online orders by the minute. Picking, out-of-stock substitution and delivery status write-back all fall inside the same order cycle, so product data and selection Q&A must work against several documents at once: the instant order, the picking list, the out-of-stock report and the delivery exception form. Material versions and order states can both change within the same hour.
Compared with store inventory counting and loss prevention, which work on a periodic stocktake cycle, this category handles online orders continuously and per order. The consequence of error is concrete: delivery timeout compensation, penalties for falsely marked delivery, and customer refunds. Answers that cite product specifications, substitution rules or exception handling must be traceable to the version of the material that the specific order relied on.
The citation tracing step therefore has to keep three things separate: what a retrieval hit returned, what the source material actually states, and what authorization state the order is in.
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 output is an array of citations, and the length may be zero. A zero-length array means the output path still runs; it does not mean the workflow reported an error and stopped.
The citation limit is measured in tokens. The minimum relevance value is a filter threshold and applies only under semantic retrieval or result reranking. Neither of these settings performs permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. Reranking requires a model and runs according to its configuration. Image retrieval requires image input and the corresponding model capability; the optional paths listed in the overall retrieval flow do not all execute on every request.
Whether a retrieved result is adopted still has to be checked against the effective version and the authorization state. Retrieval is not authorization, and a relevance score is not a category, entity or permission filter.
Hypothetical drill sample: on 10 June 2025 at 14:00, a supermarket home-delivery store handles instant order A-0610-14. A picker submits an out-of-stock report. The source text reads: "Within order A-0610-14, the 500 ml bottle of a certain brand of soda water is out of stock; substitute the same brand in 480 ml; the price difference is borne by the store; the substitution must be confirmed by the duty supervisor." A proposed sentence reads: "This order has been substituted with the same brand in the smaller size, and the price difference has been refunded." The point to be judged is that the source text only states that duty supervisor confirmation is required. It does not state that confirmation was completed, and it does not state that a refund was executed.
Configuration and implementation checks
The table below lists only settings or implementation checks that the official documents describe and that relate to this page's primary task. The values are starting points for this hypothetical trial, used for comparison against observed values from the same batch of orders. They are not defaults.
| Configuration or check item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the product data base and the out-of-stock report base | Check the vector model identifier first; exclude any base whose model differs; do not run cross-model retrieval comparisons |
| Search mode | Hybrid retrieval | Model and number questions need full-text recall; natural-language questions need vector recall; the two paths are merged through RRF |
| Minimum relevance | Hypothetical starting point 0.3 for this trial | Compare against the score distribution of retrieval results from the same batch of orders, and observe whether checkable source text remains after filtering |
| Citation limit | Hypothetical starting point 2000 tokens for this trial | Compare against the fragment lengths of the picking list and the report form, and observe whether citations are truncated |
| Result reranking | A rerank model is configured and enabled according to its settings | Compare against candidate ordering when it is not enabled, and check whether the rerank score is used for filtering |
| Citation input handoff | Retrieval output is passed item by item to the answering step | Place the answer sentence beside the cited source text and check whether subject, object and state agree |
Three common errors
- Treating a retrieval hit as completed authorization. The input is the out-of-stock report for order A-0610-14. The comparison object is the duty supervisor confirmation record. If that record is missing, the observable result is that the substitution remains pending confirmation, and the draft cannot state that it was executed.
- Treating an empty citation array as a workflow failure. The input is a retrieval that matched no fragment. The comparison object is the node output, a citation array of length zero. The observable result is that the downstream step still receives that array, and no conclusion follows that the workflow terminated.
- Treating a similarity score as a permission or entity filter. The input is a similar-product question spanning stores. The comparison object is the set of knowledge bases the current account can access. The observable result is that a higher or lower score does not change the readable record set, and records outside the account's scope still require separate verification.
How to verify the setup
- Citation traceability check. Input the hypothetical source text above and the proposed sentence. Compare three expressions: "must be confirmed by the duty supervisor" against "has been substituted" and "the price difference has been refunded." The proposed sentence is supported only when both the confirmation record and the refund execution record exist and agree on subject, order number and time. Otherwise the item is recorded as pending verification.
- Empty-result handoff check. Input the explicit paper citation array
[]. Compare it against the design description of the downstream answering step. The empty-result handoff condition is met only when the downstream step continues to execute on that array without reporting an error. No claim is made that a fabricated rule will always return an empty array. - Material currency check. Input the current version and a historical version of the product data. Compare the version identifier used by the same order. The source is traceable only when the cited fragment corresponds to the version in effect for that order. When the version identifier is missing, record the evidence gap and keep the judgment unsupported.
Across all three checks, a pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Missing evidence remains missing.