What the documents look like
Amazon FBM fulfillment differs from FBA replenishment in that the seller performs fulfillment and sets the return address, and no platform warehouse inspection step applies. The main document surface for this category consists of order picking sheets, shipment confirmation sheets, return address settings, and logistics tracking tables. Carrier selection, tracking number confirmation, and undelivered complaint handling consume much of the handover workload. When any of these items has an unclear status at handover, the outcome can include order cancellation, A-to-z claims, or a decline in store performance. These documents update per order and daily. A retrieval result supplies cited content only; it does not change the authorization state of any document.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model, and its output is an array whose length can be 0. Search mode can be semantic retrieval, full-text retrieval, or hybrid retrieval. Hybrid retrieval merges results through RRF, a rerank model reorders candidates, minimum relevance takes effect only under semantic retrieval or when result reranking is used, and the citation limit is measured in tokens. A retrieval hit does not constitute operational authorization, and a similarity score does not distinguish job role, store, or handover stage.
Hypothetical drill sample: the source text to be checked consists of three sentences. First, "Picking sheet P-0417, order number XXX-0001, status: picked, pending review." Second, "Shipment confirmation sheet S-0417, order number XXX-0001, status: draft, not submitted." Third, "Return address setting R-0417, site: North America, status is pending review, effective date not stated." The draft sentence to be judged is: "Order XXX-0001 has completed shipment confirmation, the return address is effective, and handover can proceed." That sentence merges pending review, draft, and pending review into completed, and is unsupported.
Configuration and implementation checks
The table below centers on the sample above and lists documented settings or implementation checks relevant to this feature and primary task.
| Item | Hypothetical starting point for this trial | Reason for selection and comparison method |
|---|---|---|
| Associated knowledge libraries | Associate only the same-vector-model library containing the three materials above | Check the vector model identifier first, exclude libraries with different models, and do not perform cross-model retrieval |
| Search mode | Hybrid retrieval | Compare against semantic retrieval and full-text retrieval on the same question and observe candidate differences |
| Citation limit | Hypothetical starting point of 800 tokens | Test whether materials of different lengths are truncated, and compare completeness against a smaller value |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare false recalls and misses against higher or lower values |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether subject or order number was altered |
| Result reranking | Enabled when a rerank model is available | Compare ordering before and after reranking, and use the rerank score for filtering |
Three common errors
- Treating a retrieval hit as handover authorization. Input is the draft sentence and the recalled citations; compare the status fields in the citations against the wording of the draft sentence, and observe whether the draft sentence writes pending review, draft, and pending review as completed. If the citations do not state an effective date, the sentence is unsupported.
- Associating libraries that use different vector models. Input is the vector model identifier of each library; compare whether the identifiers match, and observe whether a single retrieval mixes libraries of different models. When identifiers do not match, exclude that library and record it, without inferring missed recalls or redundant results.
- Treating minimum relevance as a permission or entity filter. Input is the minimum relevance value and the recalled results; compare whether materials outside the same store appear in the results, and observe whether adjusting the threshold changes the appearance of unrelated business records. This value filters low relevance only and does not perform isolation.
How to verify the setup
- Input the three source sentences and the draft sentence from the hypothetical sample; compare the status fields of the draft sentence against the source text, and observe whether the draft sentence is judged unsupported. If the source text does not state an effective date, the judgment remains pending verification.
- Input the same question through semantic retrieval, full-text retrieval, and hybrid retrieval; compare the candidate content from the three paths, and observe whether order numbers and status fields are correctly recalled. A first hit and complete recall are not guaranteed.
- Input an empty-result scenario; use the explicit paper input [] to check the downstream design, and observe whether the chain continues when the cited content length is 0. Do not claim that a fictional rule retrieval necessarily returns [].