What the documents look like
Multi-warehouse allocation happens at the order routing stage and determines which warehouse ships an order. The material behind product questions and selection answers comes from WMS, OMS, ERP inventory records and platform inventory interfaces. Day-to-day documents include allocation rules, inventory synchronization tables, transfer orders and out-of-stock alerts.
For retrieval-based question answering, the questions to answer are of the form "is a given product sellable at a given warehouse, and which rule selects the shipping warehouse." A usable answer depends on complete subject, warehouse, time period and status fields. The public brief notes that this material updates daily or hourly. This page does not derive batch or real-time mechanisms from that frequency; the update cadence is background only.
The category difference that matters here is that allocation occurs before dispatch, at the routing decision. A retrieval answer that names the wrong warehouse or the wrong validity window is not a shipping handover error; it is a routing error that can lead to overselling refunds, cross-region freight increases or delayed dispatch. That consequence profile is why the retrieval step must be checked against the exact fields present in the source documents.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The correct action is to check the vector model identifier and exclude any knowledge base that does not match. Cross-model retrieval is not executed in order to compare missed recalls.
Semantic retrieval, full-text retrieval and hybrid retrieval are compared against test questions. No single mode is presumed suitable for all material. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process inside hybrid retrieval, not a switch that is turned on. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation. Retrieval does not constitute authorization.
The citation limit is measured in tokens. Question optimization performs coreference resolution and question expansion. Image retrieval requires image input and model capability. Reranking requires a model and must be enabled according to settings. The node outputs citations as an array, and the length can be zero, which means the downstream path still proceeds when nothing is retrieved.
A hypothetical paper exercise sample is provided here for checking only. Source text: "East China Warehouse No. 1 is sellable for SKU-A from 2025-03-01 to 2025-03-31, and the allocation priority places East China Warehouse No. 1 before North China Warehouse No. 2; North China Warehouse No. 2 is under inventory freeze for SKU-A during the same period." Draft sentence to be judged: "SKU-A ships from North China Warehouse No. 2 with priority in March 2025." Checking subject, warehouse, period and status fields, the draft sentence and the source text take mutually exclusive values in the same fields, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases with the same vector model | Check the vector model identifier and exclude non-matching bases |
| Search mode | Start with semantic retrieval, then try full-text and hybrid | Compare hit content against the sample question |
| Citation limit | Start at a token amount covering one complete allocation rule | Compare whether long and short material is truncated |
| Minimum relevance score | Start at 0.3 | Label relevant and irrelevant samples first, then adjust |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare the original question and the optimized question for subject changes |
| Result reranking | Enable when candidates are numerous | Compare ordering before and after reranking |
The starting values above are hypothetical values for this exercise, not official defaults.
Three common errors
- Associating knowledge bases that use different vector models. Input: the vector model identifiers of two knowledge bases. Comparison: whether the two identifiers match. Observable result: when the identifiers do not match, the bases cannot be associated together.
- Treating the minimum relevance score as a permission filter. Input: an inventory record for a warehouse outside the authorized scope. Comparison: the retrieval result against the authorized scope. Observable result: a retrieval hit does not mean there is authority to execute.
- Treating a transfer order pending review as already effective. Input: a transfer order whose status field is pending review. Comparison: the status field against the wording of the draft sentence. Observable result: a draft sentence stating the order is already effective does not hold.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare the warehouse, period and status fields of the draft sentence against the source text. Observable result: the draft sentence and the source text take mutually exclusive values, so the draft sentence is judged unsupported.
- Input two knowledge bases with the same vector model identifier. Compare whether the identifiers match. Observable result: only when they match can the bases be associated together; when they do not match, record the difference and exclude the base.
- Input an explicit paper empty result []. Compare whether the downstream node still proceeds. Observable result: the citation length can be zero and the path continues. No claim is made that a query against a fictional rule must return [].