What the documents look like
Amazon FBM fulfillment places picking, dispatch and return-address control with the seller. The platform warehouse does not perform an inspection step for these orders. The working materials are order picking lists, shipment confirmation records, return address settings and carrier tracking tables. Each order and each day produce new versions, and the field values originate in separate systems: Seller Central, carrier systems, ERP orders and customer service email.
These materials carry different facts. A picking list records a picked quantity. A shipment confirmation records a dispatch status and a dispatched quantity. A return address setting records an effective scope. A tracking table records a carrier tracking number. Retrieval for document comparison has to bring the materials for the same order into one candidate set so a person can compare whether the contract or transaction materials agree. Retrieved content is candidate material only. It does not confirm order status, authorization or fulfillment completion.
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 with different vector models cannot be associated together. When an identifier does not match, check the vector model identifier first and exclude the knowledge base. Do not run a cross-model retrieval in order to compare missed recalls.
Search mode is compared across semantic retrieval, full-text retrieval and hybrid retrieval according to the test question. No single mode is assumed to suit all materials. The citation limit is measured in tokens; test with materials of different lengths to see whether content is truncated, then adjust the limit. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; label relevant and irrelevant samples before adjusting the threshold. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process and is not written as a toggle. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and is enabled according to settings. A similarity score is a relevance number and does not perform permission or entity isolation. Retrieval does not constitute authorization.
Configuration and implementation checks
The configuration below centers on a hypothetical paper exercise sample. Sample material text: a picking list states "Order A-001, SKU-77, picked quantity 12 pieces, 10 March 2025." A shipment confirmation states "Order A-001, SKU-77, dispatched quantity 10 pieces, 11 March 2025." Draft sentence to be judged: "SKU-77 of Order A-001 has been fully dispatched according to the picked quantity." The draft sentence shares the same subject, the same period and the same field as the two materials, and the dispatched quantity and the picked quantity are mutually exclusive values.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base containing the FBM materials above | Check the vector model identifier first; exclude any base whose identifier does not match |
| Search mode | Hybrid retrieval | Compare with semantic retrieval and full-text retrieval separately on order number and SKU hits |
| Citation limit | Hypothetical starting point set to a token amount covering both full material texts | Compare with a smaller limit and observe whether material is truncated |
| Minimum relevance | Hypothetical starting point set to a middle value | Compare with higher and lower values and observe whether irrelevant fragments enter |
| Question optimization | Enabled | Compare with disabled and check whether the order number and SKU are rewritten |
| Result reranking | Enabled | Compare candidate ordering with reranking off and check whether same-order material ranks earlier |
These values are hypothetical starting points for this exercise and are not official default values.
Three common errors
- Treating a retrieval hit as authorization or status confirmation. The input is the draft sentence and the two materials. Treating a retrieved shipment confirmation as direct proof that "fully dispatched" holds is treating a retrieval result as a business fact. Check method: compare the quantity in the draft sentence with the quantity in the shipment confirmation and observe whether the two agree; when they do not agree, record the item as pending verification.
- Associating knowledge bases with different vector models. The input is two knowledge base identifiers. Associating them without checking the vector model identifier is a cross-model association. Check method: check the vector model identifiers of both bases; when the identifiers do not match, exclude and record only, and do not describe the result of running the association.
- Treating a pending-review state as granted authorization. The input is a return address change request. Judging the address as already effective while the request is pending review is a state misjudgment. Check method: compare the request status field with the status required for effect; record only the current status and preserve the evidence gap.
How to verify the setup
- For the draft sentence "SKU-77 of Order A-001 has been fully dispatched according to the picked quantity," input the picking list and the shipment confirmation text, and compare the quantity field for SKU-77 in the two materials. When the dispatched quantity and the picked quantity are mutually exclusive values, the draft sentence is not supported and is recorded as pending verification; a retrieval hit is not treated as making the sentence hold.
- For the return address change request, input the request text and the status field, and compare the current request status with the condition required for effect. When the request is pending review, record it as pending review, do not judge the address as effective, and do not supply an authorization clause.
- For cross-base association, input the vector model identifiers of the two knowledge bases and compare whether they match. Only matching identifiers may be associated together; when identifiers do not match, exclude and record only, do not run a cross-model retrieval, and do not describe missed recalls or redundant results. If retrieval returns an empty array, check the downstream design against a clearly stated paper input, and do not claim that a fictional rule retrieval necessarily returns empty.