What the documents look like
Amazon FBM fulfillment places fulfillment and returns-address setup with the seller rather than a platform warehouse. For copy drafting, the working materials are the order picking list, the shipment confirmation, the returns-address setting, and the logistics tracking sheet. These documents update per order or daily and carry concrete facts such as carrier, tracking number, and returns address. The drafting task in this step is to retrieve checkable content from these documents and decide whether a proposed sentence is supported by the material. The knowledge base search node can associate one or more knowledge bases that use the same vector model, and it outputs cited content as an array whose length can be zero.
A proposed marketing or product sentence is not validated by its tone. It is validated against the retrieved source wording and its stated scope. Material that is absent from the retrieved content remains absent, and a pending request is not an approval.
Constraints at the retrieval step
Retrieval searches the whole library; a collection is only a classification layer and does not change search behavior. Search modes are semantic search, full-text search, and hybrid search. Hybrid search merges multiple result paths through RRF and usually needs a rerank model for a second ordering. The minimum relevance value takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not act as a permission or entity-isolation control. Associated knowledge bases must share the same vector model, so libraries with different vector-model identifiers cannot be associated together. That point is checked by identifier and by excluding non-matching libraries; it does not license a cross-model retrieval run to compare missed recalls. Question optimization performs coreference resolution and question expansion, adds one model call before retrieval, and requires comparing the original question with the optimized question to check whether the subject was altered.
Image retrieval is an optional path. It requires image input and the relevant model capability, and whether an image can be retrieved also depends on whether an index was built at ingestion time. Reranking requires a rerank model and must be enabled by configuration. The overall retrieval flow does not mean every optional path executes on every request.
Hypothetical paper exercise sample. Source wording: "The returns address is set to a warehouse in Bao'an District, Shenzhen, Guangdong, applicable to orders shipped from March 1, 2025 to March 31, 2025." "The carrier is an overseas local carrier, and the tracking number is entered into the logistics tracking sheet after the shipment confirmation is generated." Proposed sentence to assess: "This product supports 30-day no-reason returns, and the returns address is Shenzhen, China." The source wording states the returns address and the applicable period. It does not state a no-reason return window. The phrases "30-day" and "no-reason" are not supported by the material and remain to be verified.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases using the same vector model | The sample material must sit in knowledge bases sharing one vector model |
| Search mode | Compare semantic, full-text, and hybrid search against the test question | Run the sample proposed sentence through each mode and compare retrieved content |
| Citation limit | Unit is token; test content completeness against actual material length | This hypothetical trial starts at 512 tokens because that value can hold the two sample source sentences; compare completeness against 256 tokens |
| Minimum relevance | Takes effect only in semantic search or with result reranking | Label relevant and irrelevant sample passages first, then compare returned content after threshold changes |
| Question optimization | Used for coreference resolution and question expansion | Compare original and optimized questions and check whether "this product" was changed to a different subject |
| Result reranking | Requires a rerank model and must be enabled by configuration | Reranking yields a score from 0 to 1; filter sample candidates by that score |
Three common errors
- Treating minimum relevance as a permission or entity-isolation mechanism. The value is a filter threshold and does not exclude unrelated businesses or unauthorized records; retrieval itself is not authorization. The check is to observe whether returned content still contains passages from another subject.
- Associating knowledge bases that use different vector models. The documentation states that only knowledge bases with the same vector model can be associated. Libraries with different identifiers should be excluded and the exclusion recorded. Do not invent missed recalls or redundant results from a joint association that was never run.
- Treating RRF as a switch that can be turned on. RRF is the process that merges multiple result paths in hybrid search, not a standalone configuration item. The check is to see whether hybrid search produces both semantic and full-text recalls, then observe the merged ordering.
How to verify the setup
- Input the sample proposed sentence "This product supports 30-day no-reason returns, and the returns address is Shenzhen, China" under semantic search and compare returned citations with the sample source wording. If the returned content includes the returns address and the applicable period but not "30-day" or "no-reason," the proposed sentence is not supported. If the returned array is empty, record the empty result and check the downstream design; do not claim that a rule-based retrieval necessarily returns an empty array.
- Input "March 1, 2025 to March 31, 2025" from the sample source wording under full-text search and compare whether the returned content hits that date range. If it hits and the content within the citation limit is complete, the date field can be checked. If only part of the date is returned, adjust the citation limit and compare again.
- Input the sample proposed sentence with question optimization enabled and compare the optimized question with the original. If the optimized question replaces "this product" with another subject or adds a return window absent from the source, record that the subject was altered and do not use that optimized result for the copy assessment. If the optimized question keeps the subject and object consistent, continue comparing returned citations with the source wording.
The verification steps test the named primary task of copy drafting against retrieved source wording. They do not test identity certification, report analysis, or any other task. Where the material does not state a field, that field remains unknown.