What the documents look like
Inbound receiving and putaway handles the first book-to-physical confirmation when goods enter the warehouse. The work revolves around the purchase arrival note, the inbound acceptance note, the putaway task, and the discrepancy report. A single arrival batch often produces several documents at once, and the subject, object, date, and quantity fields across those documents must be aligned field by field. The purchase arrival note records the item names and quantities delivered by the supplier. The inbound acceptance note records what was actually received. The putaway task records the assigned storage location. The discrepancy report records the portion where the two do not match. Verifying arrival quantities, printing barcodes, and assigning storage locations are the time-consuming daily steps, and a citation must trace back to a specific field in a specific document to support later book-to-physical adjustments and supplier communication. Material updates occur per batch and per day, which affects how current a cited source is but does not change the tracing step itself.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. When nothing is retrieved, the output path still completes. The citation limit controls how many tokens of content are cited per search. The minimum relevance value is a 0-1 filter threshold, and it takes effect only under semantic retrieval or when result reranking is used. These parameters determine the scope and amount of cited content; they do not determine whether a given record is authorized to be viewed. Retrieval is the process of bringing back candidate content, while authorization decisions, approval, and database writes belong to downstream implementation. Hybrid retrieval combines semantic and full-text recall through RRF, and reranking reorders candidates by a relevance score. These are retrieval behaviors, not permission controls.
Hypothetical paper exercise sample: a warehouse receives one arrival batch. The source material reads: "Purchase arrival note PO-2024-0311 states that supplier A delivered 200 units on March 11; receiver B completed acceptance on the same day; inbound acceptance note RC-2024-0311 records 198 units actually received; discrepancy report DR-2024-0311 records a shortage of 2 units; putaway task PT-2024-0311 assigns storage location A-03." The draft sentence to be judged is: "The received quantity for this batch matches the purchase arrival note, and no discrepancy handling is required." This sentence and the 2-unit shortage recorded in the discrepancy report take mutually exclusive values on the same quantity field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge base | Select one inbound document base | Check the vector model identifier against the supplier delivery note base separately; only identical models may be associated together |
| Search mode | Hybrid retrieval | Item names and identifiers coexist; compare recall coverage against pure semantic retrieval |
| Result reranking | Enabled | When candidates are numerous, compare ranking stability using the rerank score |
| Citation limit | Hypothetical starting point of 3000 tokens for this exercise | Compare against the default display amount and observe changes in the number of cited items |
| Minimum relevance | Hypothetical starting point of 0.3 for this exercise | Compare against 0.5 and observe whether low-scoring citations are filtered out |
| Cited content output | Array, empty allowed | Compare against an empty-array input and check the downstream handling design |
The citation limit and minimum relevance are paper exercise starting points only, not official defaults. The thinking budget field belongs to model thinking configuration and is unrelated to the citation token limit.
Three common errors
- Treating minimum relevance as a permission filter. This value only filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. Check method: run the same query, compare returned entries under different relevance thresholds, and observe whether filtered items are excluded only because of low scores.
- Treating retrieval output as authorization execution. The search node outputs a citation array and does not complete approval or writes. Check method: verify whether the citation array is passed to the answering step, compare the answer content against the cited source text, and confirm no additional business action is claimed.
- Treating a pending-review material as a completed fact. The quantity judgment in the draft sentence lacks support from the discrepancy report field. Check method: compare the draft sentence against the discrepancy report source field by field and observe whether the quantity fields are mutually exclusive.
How to verify the setup
- Verification one: input the source text of the purchase arrival note, inbound acceptance note, discrepancy report, and putaway task described above, and compare the draft sentence "The received quantity for this batch matches the purchase arrival note, and no discrepancy handling is required" against the 2-unit shortage recorded in the discrepancy report. When the same quantity field holds mutually exclusive values, the draft sentence does not hold and should be marked as pending verification.
- Verification two: input an empty citation array and observe whether the downstream answering step handles it normally. Cited content length may be zero; an empty array does not indicate that the workflow has terminated with an error, and the downstream design must be checked for handling the empty-result branch.
- Verification three: check the vector model identifiers of the two knowledge bases, exclude those with different models, and record the observed association availability. Do not perform a cross-model retrieval to compare missed recalls; only check the identifiers and record the result.
The evidence gap remains: the source material does not state whether the shortage was approved for adjustment, whether the supplier confirmed it, or whether any write occurred. Those facts stay unknown and must not be inferred from the retrieval output or the draft sentence.