What the documents look like
Peak-sale warehousing and delivery scheduling centers on short-term peak capacity and the arrangement of people, vehicles, and sites. The working materials are a peak-sale forecast table, staff shift schedules, a courier support plan, and a daily exception report. In contract and transaction material comparison these materials act as sources: the forecast table supports stocking and labor commitments, the shift schedule and support plan support statements of fulfillment capability, and the exception report supports after-the-fact reconciliation. The main data sources are historical peak-sale orders, platform campaign registrations, WMS capacity, and courier pickup capacity. These materials change across versions as scheduling proceeds, so the same field can hold different values in different versions. Any comparison must lock the same period and the same field.
The forecast table, shift schedule, support plan, and exception report differ in what they can support. A forecast table expresses an expectation, a shift schedule expresses a staffing arrangement, a support plan expresses a proposed capability, and an exception report expresses an observed deviation. A citation drawn from one of these materials carries only the meaning that material can carry. Version and period are part of the citation, not decoration around it.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model. Its output is a citation array, and the array length can be zero, which means the output path still completes even when nothing is retrieved. The citation limit is controlled in tokens. The minimum relevance value takes effect only under semantic retrieval or result reranking. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a rerank model that is enabled according to its settings. Image retrieval requires image input and the corresponding model capability. These are retrieval-side conditions. They do not constitute an authorization judgment: a retrieval hit does not mean a material has been approved, and a relevance score does not perform entity or permission filtering.
Hypothetical paper exercise sample. A warehousing and delivery project team keeps a "Peak-Sale Courier Support Plan (Version 3)." The body states: "This plan takes effect from November 1, 2025, covers three warehouses in East China, uses carriers A and B, and has a combined daily pickup ceiling of 120,000 pieces." The signature date is October 28, 2025, and the status field reads "pending review by operations lead." A separate "Staff Shift Schedule (November 1 to November 3)" states that 80 temporary workers will report. The draft sentence to be judged is: "The carrier support plan has been approved, and its daily pickup capacity of 120,000 pieces can be used for November 1 scheduling." The judgment points are these: the subject is the project team and the two carriers, the object is the plan, the status is pending review, and the effective scope begins November 1. The draft sentence treats pending review as approved, which is unsupported. The 120,000-piece figure is the ceiling stated in the plan text and does not represent actually available capacity.
Configuration and implementation checks
| Configuration or check item | Value or action on this page | Basis and boundary |
|---|---|---|
| Linked knowledge bases | Select the bases holding the forecast table, shift schedule, support plan, and exception report | They must share the same vector model; bases with different vector models cannot be linked together |
| Citation limit | For this exercise, set a starting value that covers fragments from the four material types above | Measured in tokens; adjust after comparing against candidate fragment lengths |
| Minimum relevance | For this exercise, start at a lower value so numbered materials are included | Takes effect only under semantic retrieval or result reranking |
| Search mode | Use full-text or hybrid retrieval for numbered and dated fields | Proper nouns and numbers suit keyword matching |
| Result reranking | Enable according to settings and check the score | Requires a rerank model; the score is used for filtering |
| Citation output handling | Check that each citation array entry corresponds to source text | Array length can be zero; an empty output still completes the path |
Three common errors
- Treating a pending-review status as approved. In the sample, the plan status field reads "pending review by operations lead," while the draft sentence says "has been approved." Same subject, same field, mutually exclusive values; the draft sentence is unsupported.
- Treating the plan's stated daily pickup ceiling as actually available capacity. The 120,000-piece figure is a ceiling in the plan text. Actual pickup depends on vehicles, staff, and that day's orders. Without execution records, the conclusion remains to be verified.
- Treating a retrieval hit as authorization. A search returning a citation only means the material was recalled. Signing, reaching cooperation, and contract effectiveness are three distinct facts. Retrieval does not produce approval effect.
How to verify the setup
- Input the plan source text and the draft sentence from the sample, and compare whether "pending review by operations lead" and "has been approved" refer to the same subject and the same field. When the values are mutually exclusive, judge the draft sentence unsupported and record the evidence gap.
- Input the shift schedule source text and the statement "80 temporary workers will report," and compare whether the date range November 1 to November 3 covers the target scheduling date. When the dates do not cover it, the citation does not apply to the target scheduling day.
- Input an explicit paper empty array [] and check whether the downstream answer step continues through the path with an empty citation. If the path completes and no source statement is produced, the condition is met. If the downstream step generates a source conclusion from it, record that as needing correction.
Relevance scores and semantic recall do not perform permission or entity isolation. The same-vector-model requirement for linked knowledge bases does not extend to business registration sites or other external systems. The optional paths listed in the overall retrieval flow do not all execute on every request: image retrieval requires image input and model capability, and reranking requires an enabled rerank model. A pending review does not prove the absence of operating authority. A quotation validity period does not indicate a supply period. Where the source text is silent, the evidence gap stays open.