What the documents look like
FBA preparation moves inventory into a platform warehouse, so inbound discrepancies and inventory age are the main risks. The daily and weekly materials are the FBA inbound plan, carton labels, inventory reports, and compensation claims. Sources include Amazon Seller Central, FBA inventory reports, freight forwarder tracking, and ERP purchase orders. These materials carry Chinese and English fields, and the retrieval layer has to reconcile them against a specific document version.
Semantic retrieval matches across languages, so a Chinese question can reach an English knowledge point. Full-text retrieval suits exact tokens such as plan numbers, model codes, and proper nouns. A single search request does not run every available path: image retrieval requires an actual image input plus model capability, and reranking requires a rerank model that is enabled in the settings. The 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. The node outputs citations as an array, and that array may have length zero; an empty array means the search path completed, not that the workflow errored out. The minimum relevance setting is a filter threshold that takes effect only in semantic retrieval or when result reranking is used. It is not a permission boundary and does not isolate entities. Retrieval itself is not authorization.
Constraints at the citation tracing step
The citation array is the evidence chain, so each draft sentence must be checked against the source text on the same subject, the same period, and the same field. A status field reading pending review is mutually exclusive with an approved reading. Pending review does not establish that an operator lacks permission, and it does not establish a denial.
Hypothetical drill sample for this page. Source text A: "FBA inbound plan FP-2024-0312, planned inbound quantity 600 units, destination a western United States operations center, creation date 12 March 2024, status pending review." Source text B: "Carton label file states outer carton count 50, 12 units per carton." Draft sentence under review: "This inbound plan has been approved, and 600 units can be dispatched." The comparison must hold subject, period, and field constant. The source status field reads pending review, which is mutually exclusive with approved, so the draft sentence is unsupported. Pending review does not prove an absence of operating authority, and it does not constitute a denial.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this page | How to check and what it means |
|---|---|---|
| Associated knowledge bases | Only the inbound plan base and the carton label base | Confirm the vector model identifiers match; exclude any base that does not match; do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Use full-text retrieval for identifier fields; use semantic retrieval for natural-language questions |
| Citation limit | A token amount that covers both source texts | Compare against the sample source length; avoid truncating the status field |
| Minimum relevance | A middle value chosen for this drill | Compare against returned scores; it applies only in semantic retrieval or with reranking |
| Result reranking | Enabled according to the settings | Confirm the rerank model is available; make no claim when it is not enabled |
| Citation output handoff | The citation array passed to the answering step | Check each answer sentence against the cited source text; treat an empty array as zero length |
Three common errors
- Reading a pending-review status as approved. Input is the sample source status field. Compare "pending review" with "approved." The two are mutually exclusive, so the draft sentence is unsupported.
- Treating a retrieval hit as authorization. Input is the citation array from the search node. Compare the citation content with the authorization basis. Retrieval does not produce an approval or a release.
- Forcing an association across vector models. Input is the vector model identifier of each base. Compare the two identifiers. When they differ, exclude the base and record the observation. Do not compare recall gaps across models.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the status field values. When mutually exclusive values appear, record the draft sentence as unsupported and keep the evidence gap open.
- Input a paper empty array []. Compare whether the downstream design handles it as zero length. Do not claim that retrieval under a fictional rule must return an empty array.
- Input the vector model identifiers of the two bases. Compare whether they match. Only matching identifiers allow joint association. When they differ, exclude the base and record what was actually observed.