What the documents look like
Amazon FBM operations place fulfillment and the return address under the seller, without platform warehouse inspection. The daily materials are order picking lists, shipment confirmation forms, return address settings, and logistics tracking tables. These materials update per order and per day, and they feed directly into carrier selection, tracking number confirmation, and undelivered complaint handling. Knowledge base retrieval in this category has to cover order status, return address, and logistics milestones together.
The knowledge base search node outputs cited content as an array whose length may be zero. When nothing is retrieved, the output link still runs. That documented behavior matters for after-sales flows, because an empty citation array is a valid output rather than a broken branch.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The retriever is the PostgresSQL PG Vector plugin with an HNSW index, while MongoDB stores original data and an indexes field in the dataset.datas table; one data item can map to multiple vectors. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges both paths through RRF, and Rerank reorders candidate text and returns a 0-1 score. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The citation limit is measured in tokens. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and the corresponding model capability, and reranking requires a model and is enabled according to settings. These paths do not all execute on every request.
Hypothetical paper exercise sample: the source material reads, "Order A-1001, picking list generated, return address is overseas warehouse R, shipment confirmation form unsigned, logistics tracking table records carrier C has picked up." A draft sentence awaiting judgment reads, "Order A-1001 has completed shipment confirmation and the return address is in effect." The subject needed for judgment is order A-1001, the objects are the shipment confirmation form and the return address setting, and the states are picking list generated, shipment confirmation form unsigned, and carrier has picked up.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison target |
|---|---|---|
| Associated knowledge base | Associate one knowledge base using the same vector model | Compare vector model identifiers and exclude those that do not match |
| Search mode | Hybrid retrieval | Compare retrieved content against semantic retrieval and full-text retrieval separately |
| Citation limit | Hypothetical starting value of 2000 tokens | Compare completeness against picking lists and tracking tables of differing length |
| Minimum relevance | Hypothetical starting value of 0.5 | Compare scores of labeled relevant samples against irrelevant samples |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Rerank | Enabled | Compare candidate ordering against the disabled state, decided by settings |
Three common errors
- Treating a retrieval score as an authorization judgment. Minimum relevance is a filtering threshold. Semantic recall and relevance scores do not perform permission or entity isolation, and retrieval results cannot substitute for the business verification that a shipment confirmation or return address has taken effect.
- Writing a pending signature state as completed. In the sample, the shipment confirmation form is unsigned while the draft sentence states shipment confirmation is complete. This is a same-field value contradiction, and carrier pickup cannot be used to infer that shipment confirmation is complete.
- Associating knowledge bases across vector models. Knowledge bases using different vector models cannot be associated together. The vector model identifier must be checked first and non-matching bases excluded, rather than running cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the sample source material and the draft sentence, compare the draft phrase "has completed shipment confirmation" against the source phrase "shipment confirmation form unsigned," and observe whether a same-field contradiction is identified. When the condition is met, record the draft as unsupported.
- Input the sample source material and the draft sentence, compare the draft phrase "the return address is in effect" against the source phrase "return address is overseas warehouse R," and observe whether the source states an effective status. When the source does not state it, record the item as pending verification.
- Input an explicit paper empty result [], check whether the downstream design accepts the empty citation link, and observe whether the output length is zero and the link still runs. Do not claim that a fictional rule retrieval necessarily returns [].
The optional paths shown in the overall retrieval flow do not mean every request executes all of them. Image retrieval requires corresponding image input and model capability; reranking requires a model and is enabled according to settings. Knowledge bases using different vector models cannot be associated together; verify identifiers, exclude non-matching bases, and record actual observations. A pending review does not demonstrate a lack of operating authority, and signature, cooperation agreement, and contract effectiveness are three distinct facts. When the material needed for verification is absent from the accessible retrieval scope or a defined manual input, record the gap rather than supplying an assumed authorization or effective clause.