What the documents look like
Amazon FBM fulfillment is operated by the seller, who arranges delivery and sets the return address, without relying on platform warehouse inspection. The working materials for a handover are the order picking list, the shipment confirmation, the return address setting and the logistics tracking sheet. These records originate in different places: Amazon Seller Central, carrier systems, ERP orders and customer service email. Updates follow a per-order and daily rhythm, so two people looking at the same order during a shift change may be holding different versions of the same document.
Citation tracing in this category means that each material used in operations permission and handover management can be returned to a specific source. The knowledge base search node outputs a citation array, and the length of that array can be zero. An empty array does not mean the workflow has errored and stopped. Whether cited content is actually used by a downstream answer step has to be checked line by line against the cited original text; a fluent answer is not a substitute for source evidence.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model. Knowledge bases built on different vector models cannot be associated together. During a check, compare the knowledge base identifier and the vector model identifier first and exclude those that do not match, rather than running a cross-model search to compare missed recalls.
The citation limit is controlled in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. Minimum relevance is a filter threshold, not a final ordering rule. Once question optimization, hybrid retrieval, image retrieval or result reordering is enabled, the final result may come from a fusion of several recall paths and does not necessarily follow a single vector similarity order. RRF is the fusion process for hybrid retrieval results and is not described as a standalone switch.
Semantic recall and relevance scores do not carry permission filtering or entity isolation duties. Retrieving a record does not mean the current operator holds business permission over that record. Forms collect field values; business verification, approval, database writes and external synchronization each need their own downstream implementation.
Hypothetical paper exercise: the original text of an FBM order picking list reads "Order number FBM-2041, picker A, status: picked; shipment confirmation not yet generated." A draft handover sentence reads "Order FBM-2041 has completed shipment confirmation and can be handed to the next shift." Here the subject is order FBM-2041, the objects are the picking and shipment confirmation materials, the status is picked with no shipment confirmation generated, and the draft sentence claims a completed shipment confirmation that the original text does not support. This sample is used for paper comparison only and does not represent a real order or any business action that has taken place.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only knowledge bases sharing the same vector model are selected | Compare knowledge base identifiers and vector model identifiers, and exclude bases with a different vector model |
| Citation limit | Hypothetical starting point of 2000 tokens | Compare against the length of a single handover material to see whether one full picking list fits |
| Minimum relevance | Hypothetical starting point of 0.3 | Compare against the number of recalls without filtering, and observe whether unrelated fragments are removed |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone, and observe whether numbered fields are hit |
| Result reordering | Enabled according to the setting | Compare against ordering without reordering, and confirm whether the reorder score is used for filtering |
| Citation output handoff | Answer checked line by line against cited original text | Check whether the answer cites material that does not appear in the retrieval result |
The values in the table are hypothetical starting points for this exercise, used to explain the comparison method, and are not official default values.
Three common errors
- Treating a retrieval hit as operating authorization. The input is a citation array returned by one search; the comparison object is the current operator's handover permission record; the observable result is that the retrieval result contains a certain order's material while the permission record holds no handover authorization for that order. These are different fields, and a retrieval hit does not constitute authorization.
- Writing a pending review status as already effective. The input is a draft handover sentence; the comparison object is the status field in the original material; the observable result is that the status claimed in the draft differs from the status value in the original text. A pending review does not demonstrate a lack of operating permission and does not equal a completed approval.
- Treating an empty citation array as a process failure. The input is the citation array output by the search node; the comparison object is the downstream handoff design; the observable result is an array length of zero while the downstream still continues along the empty-array path. An empty array does not equal a workflow error stop.
How to verify the setup
- Input the picking list original text and the draft sentence from the hypothetical sample, and compare the shipment confirmation status claimed in the draft against the status field in the original text. If the two values differ, the draft sentence is not supported by the original text and should be recorded as pending verification.
- Input a citation array explicitly marked as empty, and compare whether the downstream answer step continues along the empty-array path. If the downstream does not error and carries on along the empty path, the empty-citation handoff design holds. This check does not claim that a search for an invented rule will necessarily return an empty array.
- Input two knowledge base identifiers marked as using different vector models, and compare whether their vector model identifiers match. If they do not match, exclude joint association, record only the identifier comparison result and the actual observation, and make no assertion about the runtime result of a cross-model search.
This page addresses only Amazon FBM fulfillment / operations permission and handover management / citation tracing. During an operations handover, which materials, actions and requests sit in which authorization state cannot be resolved by treating retrieval or a form as authorization execution. Materials must sit within a retrieval scope the operator is entitled to access, or within an explicit manual input; when they are missing, record the gap.