What the documents look like
Amazon FBM fulfillment places the seller in charge of delivery and of the return address, without relying on platform warehouse inspection. The working material comes from Amazon Seller Central, carrier systems, ERP orders and customer service email, and the everyday documents are order picking lists, shipment confirmation sheets, return address settings and logistics tracking tables. These records carry order numbers, carrier names, tracking numbers, return addresses and multilingual customer correspondence, and they update per order and daily.
The shape of the material decides how retrieval behaves. Tracking numbers and order numbers are identifier-style content, so full-text retrieval depends on keyword hits, while natural-language questions inside customer service email depend on semantic retrieval. The knowledge base search node can associate one or more knowledge bases that share the same vector model; bases built on different vector models cannot be associated together, so the vector model identifier has to be checked first and non-matching bases excluded.
Constraints at the retrieval step
This page addresses only knowledge base retrieval and recall inside cross-border material and multilingual search. Retrieval and recall carry no authorization function, and semantic similarity together with the minimum relevance value acts as a relevance filter rather than a permission or entity-isolation mechanism. The requirement that associated knowledge bases share one vector model does not extend to external systems such as carrier portals or marketplace back offices.
The node output is an array of references, and its length can be zero. When nothing is retrieved, that output path still completes, so a downstream design that expects references has to accept an empty array rather than assume a hit.
Hypothetical paper exercise sample: the source material reads, "Order A-1042, carrier is an overseas local carrier, the tracking number field has been filled in, the return address is set to a US warehouse address, status is pending shipment confirmation." The draft sentence under judgement reads, "This order has completed shipment confirmation and the tracking number is valid." The source records only that the field has been filled in and that the status is pending shipment confirmation. It does not record that shipment confirmation has been completed, and it does not record any determination that the tracking number is valid. The draft sentence is therefore unsupported, and the evidence gap stays open.
Configuration and implementation checks
| Configuration item | Treatment on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select only one or more bases sharing the same vector model | Check the vector model identifier first, exclude non-matching bases, and do not run a cross-model retrieval comparison |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Identifier content uses full-text or hybrid; customer service questions use semantic or hybrid |
| Reference limit | Unit is tokens; test completeness with source material of differing lengths | A hypothetical starting value of a few hundred tokens for this exercise, chosen because of the sample paragraph length, compared against shorter and longer values |
| Minimum relevance | Takes effect only in semantic retrieval or when result reranking is used | Label relevant and irrelevant samples before adjusting the threshold; not a permission filter |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a ReRank model and must be enabled in settings | Performs a second ordering of candidate text; the 0-1 score can be used for filtering |
RRF is the fusion step for hybrid retrieval results and is not a switch. Image retrieval requires image input together with the corresponding model capability; when no image is supplied, that path does not execute.
Three common errors
- Treating a retrieval hit as business authorization. Input is the retrieval question for sample order A-1042; compare the retrieval output against the shipment confirmation status. Writing the recalled content directly as "shipment confirmation completed" confuses retrieval with authorization and must be recorded as unsupported.
- Treating a pending review state as approved. Input is the status "pending shipment confirmation"; compare it against the draft sentence "shipment confirmation completed". Both share the same subject and the same field but hold mutually exclusive values, so the draft is unsupported and cannot be written as fact.
- Associating knowledge bases across different vector models. Input is two bases with different vector model identifiers; compare whether the identifiers match. Forcing the association violates the same-model requirement, so the non-matching base is excluded and the actual observation is recorded.
How to verify the setup
- Input the hypothetical sample source text together with the draft sentence "This order has completed shipment confirmation and the tracking number is valid". Compare whether the source records completion of shipment confirmation and any determination of tracking number validity. The expected result is unsupported, with the evidence gap preserved.
- Input the tracking number query for sample order A-1042. Compare the recalled content from semantic retrieval against full-text retrieval and observe whether the identifier-style content is hit. A failure branch is recorded as a miss, and a first-position hit is not guaranteed.
- Input an explicit paper empty result []. Check whether the downstream design accepts the empty reference path, without claiming that a fictional rule retrieval necessarily returns []. At the same time, check the vector model identifier and exclude non-matching bases.