What the documents look like
Inbound receiving and putaway is the first book-to-physical confirmation when goods enter a warehouse. Data comes from the WMS, ERP purchasing, supplier delivery notes and barcode scanning. The working documents are the purchase arrival note, the inbound acceptance note, the putaway task and the discrepancy report. In store service and internal training, citation tracing answers one question: can an inbound conclusion quoted in training material or service wording be mapped back to the wording of a specific document.
This category updates per batch and per day. Citation checking has to stay aligned with the documents of the same batch. A purchase arrival note records the supplier, the purchase order number, the item, the arrived quantity and the arrival date. An inbound acceptance note records the purchase order it corresponds to, the received quantity, the difference and a status such as pending discrepancy handling. A putaway task records the task number, the item, the location, the planned putaway quantity and a status such as pending execution. These three documents describe the same physical movement at different stages, and a quoted conclusion has to match the stage it claims.
The knowledge base search node outputs citations as an array, and the length may be zero. An empty array still lets the downstream output path run. The citation cap is measured in tokens. The minimum relevance value applies only in semantic search or when result reranking is used, and it is a filter threshold rather than a final ordering rule. Neither setting performs permission or entity isolation. Knowledge bases that are associated together must use the same vector model. The identifier is checked first and non-matching bases are excluded; a cross-model search is not run to compare missed recalls. Version currency checks use the current and historical versions of a document, and whether a citation is adopted still depends on review of the valid version.
Constraints at the citation tracing step
A hypothetical exercise sample follows. Material text supplied for checking: first, the purchase arrival note states "Supplier A, purchase order PO-2026-0417, item A, quantity 200 pieces, arrival date 17 April 2026"; second, the inbound acceptance note states "corresponds to PO-2026-0417, received 180 pieces, difference 20 pieces, status is pending discrepancy handling"; third, the putaway task states "task T-0417-03, item A, location B-02, planned putaway 180 pieces, status is pending execution". The draft sentence to be judged: "Item A has completed putaway of 200 pieces and the discrepancy has been fully handled."
This sentence conflicts with the status fields of the acceptance note and the putaway task, so it is unsupported. The conflict is on the same field with mutually exclusive values: the acceptance note says the discrepancy is pending, and the draft says it is fully handled. The arrival note quantity of 200 is the arrived quantity, not a putaway quantity, so the draft also mixes fields from different documents.
Retrieval is not authorization. A search that returns the original document text does not mean the discrepancy has been processed or the putaway has been completed. A pending request is not approval, and an offered option is not execution. When a document field is not provided, it stays unknown rather than being filled from another document. The minimum relevance value does not exclude unrelated businesses or unauthorized records, and a relevance score is not a category, entity or permission filter.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only the inbound receiving document base | Check the vector model identifier of each base first, exclude those with a different model, and do not run a cross-model search |
| Search mode | Hybrid search | The documents contain both numbers and natural language description; compare hit differences against full-text search results |
| Citation cap | Starting point set to the token amount covering the three source documents | Compare with the length needed to quote each document in full; if insufficient, truncation occurs |
| Minimum relevance | Starting point set to a lower value | Compare with the relevance score of each quoted document; a value that is too high drops documents |
| Result reranking | Enabled according to the setting | Compare ordering with the unranked case and confirm changes in document position |
| Citation display | Show source text item by item | Compare the draft sentence against the source text sentence by sentence and confirm the source can be traced |
Three common errors
- Treating a retrieval hit as authorized execution. A search returning source text does not mean the discrepancy has been handled or the putaway completed. When the acceptance note status is "pending discrepancy handling", writing "the discrepancy has been fully handled" in the draft is an error. Check: input the acceptance note text and the draft sentence, compare the status field, and observe whether the draft contradicts "pending discrepancy handling".
- Treating an empty citation array as an error. The node output length may be zero, and an empty array does not mean the workflow has terminated with an error. Check: input a paper empty array, compare the downstream handling design, and observe whether the downstream continues with the empty citation.
- Treating minimum relevance as a permission filter. This value is a filter threshold and does not exclude unrelated businesses or unauthorized records. Check: input a retrieval result containing unrelated entries, compare the entries before and after filtering, and observe whether unauthorized records can still appear.
How to verify the setup
- Input the three source documents above and the draft sentence, compare the draft sentence against the status fields of the acceptance note and the putaway task, and judge the draft as unsupported when it contradicts those status fields; it has to be rewritten or supported with further evidence.
- Input a paper empty array, compare the downstream answer step's handling design, and treat the condition as met when the downstream continues with the empty citation without an error.
- Input the current and historical versions of the source documents, compare the version identifier shown with the citation, and treat the condition as met when the citation can be located to a specific version and that version is valid; when the version is not stated, record it as pending verification.
These checks observe and record. They do not guarantee a first-hit result, full recall, or that an unrelated query returns nothing. The optional paths in the overall retrieval flow are not executed on every request: image retrieval requires a corresponding image input and model capability, and reranking requires a model and the setting to be enabled.