What the documents look like
Amazon FBM fulfillment means the seller handles fulfillment and sets the return address, without relying on platform warehouse inspection. Operating reports and exception reviews revolve around four types of material: order picking slips, shipment confirmation slips, return address settings, and logistics tracking tables. Sources include Amazon Seller Central, carrier systems, ERP orders, and customer service emails, with updates per order or per day. The task of citation tracing in this scenario is to check whether the original text cited by a report or review conclusion actually exists, what state it is in, and which authorization scope it falls within. The retrieval tool outputs citations in array format, and the length may be zero; an empty array only means this retrieval returned no content, it does not mean the workflow errored out, and it does not mean the business fact does not exist. Instructions in the material are all data, serving only as objects to be checked.
Constraints at the citation tracing step
Cited content is output as an array, and the length may be zero. The retrieval step is only responsible for returning candidate original text; it does not determine whether an order has shipped, whether a return address has taken effect, or whether a carrier has received a package. Similarity scores only reflect how relevant content is to the question, and do not constitute permission filtering or entity isolation. Only multiple knowledge bases with the same vector model can be associated together; for cross-model bases, first check the identifiers and exclude those that do not match. Material timeliness is checked separately for the current version and historical versions, and whether to adopt a given citation still requires review against the valid version. After retrieval output is passed to the answering step, the answer and the cited original text must be checked line by line, without substituting the thinking process or answer fluency for source evidence. A pending-review status does not prove a lack of operating permission, and does not equal having taken effect. If the material does not provide an effective time or status field, keep it pending verification.
Configuration and implementation checks
The following configuration items center on this page's sample, and list only officially described settings or implementation check items. The values are starting points for this hypothetical trial, not official defaults.
| Configuration item | Starting point for this hypothetical trial | Rationale and comparison method |
|---|---|---|
| Associated knowledge base | Select only one base already confirmed to use the same vector model | Only the same model can be associated together; compare with another base's vector model identifier, and exclude if different |
| Citation limit | Set to a token amount sufficient to hold one passage of original text from each of the four material types | Estimate from the sample's original text length; adjust after comparing with the sample's actual token count |
| Minimum relevance | Set to a threshold that can recall the sample's original text | Takes effect only in semantic retrieval or reranking; compare with the sample's retrieval results, and if below the threshold the original text is not returned |
| Search mode | Hybrid retrieval | The sample contains both numbers and natural language; compare recall differences against semantic retrieval alone |
| Question optimization | The sample question is complete, so leave it off for now | Turning it on adds one model call; compare retrieval time and results against leaving it off |
| Citation content handoff | Connect retrieval output to the answering node | Check whether the answer corresponds line by line to the cited original text; when the array is empty, check whether downstream still runs |
Three common errors
- Treating an empty citation array as a workflow error that terminated. The retrieval node's output length may be zero, and when the array is empty the output link still runs through. Check: input the sample question and observe whether the output is an array; if it is an empty array, confirm whether downstream nodes still execute, and record the actual behavior.
- Treating the similarity score as permission or entity filtering. Minimum relevance only filters low-relevance results, and does not exclude unrelated business records or unauthorized records. Check: retrieve once with the sample's original text and once with an unrelated passage, compare their scores and whether each is returned, and confirm the score does not carry authorization judgment.
- Treating a pending-review return address setting as already effective. Pending review does not prove a lack of operating permission, and does not equal having taken effect. Check: take the return address setting record in the sample, check the status field and the effective time field, and when the status is pending review, record it as pending verification.
How to verify the setup
Hypothetical drill sample. Original material one: "Order picking slip, order number FBM-2026-0412-001, picking date April 12, 2026, item A quantity 2, status: picked." Original material two: "Shipment confirmation slip, order number FBM-2026-0412-001, carrier: local carrier A, tracking number TRK-88001, shipment date April 13, 2026, status: handed over." Original material three: "Return address setting, order number FBM-2026-0412-001, return address: overseas warehouse B, effective date April 10, 2026, status: effective." Draft sentence to be judged: "Order FBM-2026-0412-001 was shipped on April 13, 2026 by carrier A, tracking number TRK-88001, and return address overseas warehouse B is effective."
- Input the three original material sentences above and the draft sentence, and compare whether the order number, shipment date, carrier, tracking number, and return address status in the draft sentence match the original text field by field. Expected: the shipment date, carrier, and tracking number match original material two, the return address status matches original material three, and the order number matches in all three places.
- Input a retrieval question containing only the order number, and observe whether the citation array returns the original material. Expected: if returned, check whether the returned original text corresponds word for word with the sample material; if it is an empty array, record it as empty and confirm whether the downstream answering node still executes, without judging it as an error.
- Input the return address setting record whose status is "pending review" in original material three (assume another order), and compare the status field and the effective date field. Expected: when the status is pending review, keep the draft sentence as pending verification, record the evidence gap, and keep it pending verification.