What the documents look like
FBA preparation moves inventory into platform warehouses, so operating reviews and exception reviews center on inbound receipt and inventory age risk. The working materials are the FBA inbound plan, carton labels, inventory reports and compensation claims. Their evidence sits across Amazon Seller Central, FBA inventory reports, freight forwarder tracking and ERP purchase orders. Citation tracing means that each conclusion in a review points back to specific source text, with the version and validity scope of that material marked. Updates arrive daily and weekly, so quantities and statuses in a report may already have changed between two updates; the extraction time must be recorded with any citation.
A paper exercise sample makes the tracing requirement concrete. Source text one: the FBA inbound plan dated 10 March 2025, version A, states that the replenishment batch from the Shenzhen warehouse to ONT8 is 600 units, with a planned arrival date of 25 March 2025. Source text two: the freight forwarder tracking record of the same date states that the batch has been handed to the carrier and is in transit, with no warehouse receipt shown. Source text three: the inventory report dated 28 March 2025 shows 420 sellable units at ONT8. The draft sentence under review reads: the batch completed warehouse receipt on 25 March and inventory shows no discrepancy. The comparison result is that the source text supports only handover to the carrier and in-transit status; completed receipt and inventory discrepancy are not supported, so the draft sentence does not match the materials.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array means the output path still runs; it does not by itself establish that the workflow terminated with an error. The retrieval output must be passed to the answering step, and the answer must then be checked line by line against the cited source text. Similarity and minimum relevance act only as filtering thresholds; they do not perform permission or entity isolation, and retrieval is not authorization.
In this category, inbound plans, carton labels, inventory reports and compensation claims carry different effective scopes. A planned arrival date is not a completed receipt. A tracking record showing carrier handover is not warehouse confirmation. A later inventory report shows a quantity at a point in time and does not by itself reconcile an inbound batch. Pending review is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Business verification, approval, database writes and external synchronization each require their own downstream implementation.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the bases holding the inbound plan, carton labels, inventory reports and compensation claims | Check the vector model identifier of each base first and exclude those that differ; no cross-model retrieval comparison is performed |
| Citation limit | Hypothetical starting point of 800 tokens | This is a starting value for this exercise, used to observe how the number of cited items changes, compared with prior report citation lengths |
| Minimum relevance | Hypothetical starting point of 0.35 | A starting value for this exercise, compared with the candidate set when filtering is not applied, to observe missed and excess citations |
| Search mode | Hybrid retrieval | Exact queries such as batch numbers coexist with natural-language review questions |
| Result reranking | ReRank enabled according to the setting | Secondary ordering when candidates are numerous; the rerank score is used for filtering |
| Citation display check | Compare each answer line against the cited source text | Source text is the evidence; answer fluency does not replace the source |
The configuration popup shows only options the current model supports. Options that are not displayed remain unknown, and no default value, limit or unit should be assumed for them. The thinking budget is a model reasoning budget and is not interchangeable with the citation token limit. RRF is a fusion process for combining result paths, not a toggle to be switched on. Image retrieval requires an image input and the corresponding model capability; reranking requires an available model and the setting to be enabled. Optional paths in the overall retrieval flow are not executed on every request.
Three common errors
- Treating a retrieval hit as authorization or completed approval. Retrieval returns citation content only; warehouse confirmation, compensation approval, database writes and external synchronization each need their stated downstream implementation. A pending request is not approval, and an unsigned termination is not a completed termination.
- Associating knowledge bases that use different vector models. The documentation requires the same vector model for joint association. The correct action is to check the identifiers and exclude those that do not match, and to record what is observed rather than running a cross-model retrieval to compare missed recall.
- Treating the thinking budget as the citation limit. The thinking budget is a model reasoning budget, while the citation limit is measured in tokens. The two cannot be substituted for each other, and configuration items not shown in the popup remain unknown.
How to verify the setup
- Input the three source sentences and the draft sentence from the paper exercise sample above. Compare whether the phrases completed warehouse receipt and no discrepancy in the draft sentence have support from source text with the same subject, the same period and the same field. The source shows only in-transit status, and the inventory report date is later than the planned arrival date, so those phrases are unsupported. If the draft sentence is changed to handed to the carrier and in transit, it matches the source text.
- Input a query containing only the batch number ONT8-0310. Compare the cited items returned by hybrid retrieval against those returned by semantic retrieval alone. If the full-text path returns source text from the carton labels or the inbound plan, the condition for exact-number retrievability is met. If neither path returns a result, record the empty array and do not infer a system error.
- Use the paper input of an empty citation array to check the design of the downstream answering step, and observe whether that path continues to execute. If the downstream step outputs no source retrieved for the empty citation, the empty-result handover holds. If the downstream step halts, record it as an implementation gap to be verified, without asserting that retrieval necessarily returns an empty result under any invented rule.
Across all three checks, a judgment requires the same subject, the same period and the same field. Only mutually exclusive values in the same field constitute a contradiction. Statements that are not recorded, or that show only a related stage, remain unsupported or pending verification. Materials needed for a check must sit within an accessible target retrieval scope or in a clearly specified manual input; when they are missing, record the gap.