What the documents look like
FBA preparation moves inventory into a platform warehouse, so the main risks sit at inbound receiving and inventory age. The working materials are inbound shipment plans, carton labels, inventory reports and reimbursement claims. They arrive from Amazon Seller Central, FBA inventory reports, freight forwarder tracking and ERP purchase orders. Updates follow a daily or weekly rhythm.
These materials mix Chinese, English and other languages in the same file set. Shipment identifiers, SKUs, shipment numbers and natural-language notes sit side by side. A retrieval step therefore has to serve two different question shapes at once: a semantic question written in one language that should reach a passage written in another, and an exact identifier query where the string itself is the target. Retrieval output is quoted content passed downstream. It does not decide whether a person may act on that content.
Constraints at the retrieval step
The knowledge base search node can attach one or more knowledge bases that share the same vector model. The smallest search unit is the knowledge base. A collection inside a base is a classification container and does not change search behaviour.
Semantic retrieval compares vector distance and handles close meaning and cross-language matching. Full-text retrieval suits identifiers, model numbers and proper nouns. Hybrid retrieval runs both paths and merges them through RRF. Result reordering uses a ReRank model and produces a 0-1 score. The reference cap is measured in tokens, not in a fixed number of items. The minimum relevance value applies only under semantic retrieval or when result reordering is used. Question optimization performs reference resolution and question expansion before retrieval. The reference output is an array and its length can be zero, so the downstream branch still runs when nothing is found.
Similarity and relevance scores are not a category filter, an entity filter or a permission filter. They do not separate unrelated businesses and they do not exclude records a user is not allowed to see. Retrieval is not authorization. Bases built on different vector models cannot be attached together; the vector model identifier is checked first and non-matching bases are excluded rather than searched for comparison.
Hypothetical paper exercise. The source text under review reads: "This inbound shipment plan is numbered FBA-2024-0315 and covers shipments sent to the United States site between 15 March 2024 and 31 March 2024. The carton label file records an outer carton count of 120 and shipment number SHIP-7788. The inventory report as of 20 March 2024 shows sellable inventory for this SKU at 0." A proposed sentence reads: "This SKU had sellable inventory on 20 March 2024, so no replenishment is needed." The source states the field value as 0 on that date and the proposed sentence states the opposite value for the same field, same subject and same period, so the proposed sentence is not supported. A separate handover request reads: "Please arrange replenishment under FBA-2024-0315 and generate the purchase order." This asks for a business action. Retrieval and a form do not grant that authority, so the request is recorded as pending verification and requires a separate approval and downstream implementation.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Attached knowledge bases | Attach only FBA material bases sharing one vector model | Check the vector model identifier first, exclude non-matching bases |
| Search mode | Try semantic and full-text retrieval separately | Compare an identifier question under full-text against a natural-language question under semantic |
| Reference cap | Start from the actual length of the sample material | Compare content completeness across short and long chunks |
| Minimum relevance | Mark relevant and irrelevant samples first, then pick a value | Applies only under semantic retrieval or reordering; compare before and after filtering |
| Question optimization | Enable for the multi-turn follow-up sample and compare | Check whether the optimized question keeps the correct subject |
| Result reordering | Compare once a model is present and the setting is enabled | Compare candidate ordering before and after reordering |
Three common errors
- Treating a retrieval hit as authorization to act. Input: "Please arrange replenishment under FBA-2024-0315." Retrieval returns the inbound shipment plan text. The comparison point is whether that text carries an approval or execution authority field. The observed result is that the source does not state such authority, so the item stays pending verification.
- Treating the minimum relevance value as a permission or entity filter. Input: a query reaching across business entities. The comparison point is whether returned content carries a permission field. The observed result is that a relevance score does not perform isolation, so the access scope needs separate checking.
- Treating an empty result as a negative conclusion. Input: an explicit paper input of
[]. The comparison point is whether the downstream branch is designed for an empty array. The observed result is that the branch runs with empty content, and this does not establish that the material is absent.
How to verify the setup
- Input the hypothetical source text above together with the proposed sentence "This SKU had sellable inventory on 20 March 2024." Compare the inventory report field value on the same subject and the same date. The observed result is 0 against the opposite value in the proposed sentence, so the sentence is not supported.
- Input the handover request "Please arrange replenishment under FBA-2024-0315 and generate the purchase order." Compare whether the material contains an approval or authority field. The observed result is that no such field is stated, so the item remains pending verification and replenishment is not executed.
- Input the identifier question "Which inbound shipment plan does SHIP-7788 belong to?" Compare the full-text and semantic retrieval returns. The observed result is whatever content is actually matched; first-position hits and complete recall are not guaranteed. If an empty array is returned, check the downstream design against that empty result.
These checks describe expected observations from a paper exercise. They do not record a completed test, an approval, a deployment or a refund.