What the documents look like
Parcel handoff happens at the point where a package leaves the warehouse, and the carrier scan carries responsibility. For procurement and supplier records, citation tracing works on four traceable materials: electronic waybills, handoff lists, pickup scan records and exception item registers. These materials come from the courier system, WMS shipment records and carrier bills, and they update per batch and per day.
The category difference matters here. Warehouse packing concerns internal operations; courier handoff concerns carrier scan responsibility. A handoff list can record that a shipment group transferred 50 parcels to a carrier outlet and that the parcels show a scanned pickup status. That record does not by itself establish that supplier records were verified, that a bill was reconciled or that any approval took effect. Citation tracing keeps these facts separate.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its citation output is an array, and the length may be zero, so the downstream path still runs when nothing is retrieved. The citation upper limit is measured in tokens. The minimum relevance value applies only under semantic retrieval or result reranking. Retrieval is a recall process; it does not perform authorization checks. Material version and subject scope must be verifiable in the source text.
A paper exercise sample is set out here for checking. Source text: "On 10 March 2025, the shipment group handed 50 parcels, waybill numbers SF-0001 through SF-0050, to the carrier outlet; handoff status is scanned pickup." A second source line reads: "The exception register for this batch lists zero entries." A draft sentence under review reads: "This batch has completed pickup handoff, and the shipment quantity in the supplier records matches the bill." The question is whether the draft sentence is supported by the source text. It is not. The source records a scanned pickup status only; it contains no supplier record field and no bill comparison field.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Check selected knowledge bases and material versions; only knowledge bases with the same vector model can be associated together | The sample materials must sit inside one searchable scope |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval, chosen by material characteristics | Waybill numbers and dates suit full-text retrieval checks |
| Citation upper limit | Citation content length is controlled in tokens | For this exercise a starting value of 500 tokens is chosen because the sample source text is short; compare against the model context and adjust |
| Minimum relevance | Applies only under semantic retrieval or result reranking | Filters low-relevance fragments; does not isolate permissions |
| Citation output | Output is an array, and length may be zero | An empty array still allows the downstream path to run; verify separately |
| Result reranking | Requires a ReRank model and must be enabled in settings | Secondary ordering when candidates are numerous; does not guarantee a first hit |
The search node's documented behaviour, the configurable fields above, manual comparison work and business actions that need separate system implementation are four different things. RRF is a fusion step for mixed results, not a switch. Thinking budget is a reasoning budget and is not a citation token limit.
Three common errors
- Treating retrieval recall as authorization confirmation. In the sample, the handoff list records only scanned pickup status. Retrieving that material does not mean supplier records were approved or that the bill was reconciled.
- Treating a pending state as effective. If the draft sentence cites an unsigned supplier record, that record sits at a pending review stage and does not constitute an effective contract or a confirmed quantity.
- Forcing knowledge bases with different vector models into one association. Cross-model association does not meet the documented requirement. Check identifiers, exclude non-conforming knowledge bases and record what is observed. Do not run cross-model retrieval to compare missed recall.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the draft phrase "completed pickup handoff" against the source phrase "scanned pickup" for the same subject, the same period and the same field. The source records scan status only and states nothing about supplier records matching the bill, so that part of the draft sentence is unsupported.
- Input the sample source text and a request asking whether the batch has received supplier confirmation. Compare whether the source contains a supplier confirmation field. The source has no such field, so record an evidence gap. Do not classify this as a denial or a conflict.
- Input an explicit empty citation array [] on paper and check whether the downstream answer step still runs. This check verifies the empty-result handling design only. It does not claim that a fictional rule retrieval always returns [].
For each check, state the input, the comparison and the observable result. If the material is not inside an accessible search scope and no manual input exists, record the gap and leave it open. A pending request is not approval, an offered option is not execution and an unsigned termination is not a completed termination.