What the documents look like
Multi-warehouse allocation decides the shipping warehouse during order routing, so inventory procurement and supplier records touch WMS, OMS, ERP inventory and platform inventory interfaces at the same time. Split-warehouse rules, inventory synchronization tables, transfer orders and shortage alerts are the daily materials, and the brief gives a daily or hourly update frequency. Citation tracing answers a narrower question: which material, which version and which passage each conclusion in procurement and supplier records comes from, and whether that material is in a state that can be cited at all. The retrieval step only recalls candidate source text and returns a citation array; authorization, approval and writes belong to downstream implementation.
Constraints at the citation tracing step
The dataset search node can associate one or more knowledge bases that use the same vector model, and its output is a citation array whose length may be zero. When nothing is found, the output path still runs. The citation limit is controlled in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold. Hybrid retrieval merges through RRF, reranking requires a model and must be enabled according to settings, and image retrieval requires image input plus the corresponding model capability. Retrieval does not carry permission or entity isolation, and the minimum relevance value cannot serve as an admission filter for suppliers or warehouses.
Hypothetical paper test case: the source material reads, "The March 1, 2025 version of the East China Region Split-Warehouse Procurement Replenishment Note states: replenishment requests for the Suzhou warehouse are submitted by the East China supply chain planning role and take effect after review by the regional procurement lead. The note applies to self-operated warehouses in East China and is valid through June 30, 2025." The draft sentence to be judged is, "The Suzhou warehouse replenishment request has been approved and can be issued directly to the supplier." That sentence turns "submitted and subject to review" into "has been approved," and no review completion record is provided.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this test | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only procurement and supplier record bases that share one vector model | Compare against the base identifier list and exclude any base with a different vector model |
| Search mode | Hybrid retrieval | Compare recall of identifier-type queries against semantic retrieval alone |
| Minimum relevance | Starting point 0.3 | Compare with 0.5 and observe whether irrelevant fragments decrease |
| Citation limit | Starting point 3000 tokens | Compare with 1500 and observe whether cited source text is truncated |
| Result reranking | Enabled | Compare candidate ordering with reranking off, after confirming the model is available |
| Citation output handling | Check whether the array is passed to the answering step | Compare with the empty array [] branch to confirm the downstream design |
The starting values are comparison baselines for this page's hypothetical test, not official defaults.
Three common errors
- Treating a retrieval hit as authorization. The error is judging a replenishment request approved on the strength of a recalled split-warehouse rule. Check: input the source material above and the draft sentence, compare "submitted and subject to review" with "has been approved," and observe whether any review completion record appears; with no record, the claim is unsupported.
- Treating the minimum relevance value as a supplier filter. The error is assuming that raising the threshold excludes unauthorized suppliers. Check: input the same query, compare the citation arrays before and after the threshold change, and observe whether only the number of recalled items changes while the permission judgment still lacks a basis.
- Associating bases that use different vector models. The error is comparing recall differences between the two bases directly. Check: input the base identifiers and vector model identifiers, compare whether they match, and exclude mismatched bases rather than running a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the source material above and the draft sentence, compare whether the draft restates the source with the same subject, the same period and the same field, and observe whether "has been approved" is supported by the source. If it is unsupported, mark it as pending verification and keep the review record gap open.
- Input one explicit query, compare the citation array length with the citation limit, and observe whether truncation or an empty array [] is handled as documented. Do not assume the first hit or full recall.
- Input the base identifier list, compare the vector model identifier of each base, and observe whether any mismatched item exists. Exclude mismatches and record the actual observation without inferring the runtime result of cross-model association.