What the documents look like
FBA preparation moves inventory ahead of demand into platform warehouses, so the operating risk sits with inbound receiving and inventory age. The document trail for this category runs through the inbound plan, carton labels, inventory reports and compensation claims. Conversation logs and audit work on a different layer: they take the session records grouped by chatId under an application, plus the single-turn run detail located by dataId, and return fields and statuses that can be checked against the source material. They do not decide replenishment quantities or inbound differences themselves.
Business data arrives from Amazon Seller Central, FBA inventory reports, freight forwarder tracking and ERP purchase orders. Whether any of that business data reaches a conversation record depends on the actual orchestration and the response the interface returns. A log entry existing does not mean the related business data was fully retained. For operating reviews this matters because a reviewer may treat a stored conversation as a complete record of an inbound decision, when the stored content only reflects what the orchestration passed through.
Constraints at the conversation logs step
Session listing is pulled by appId, offset, pageSize and source. When source is set to api, the listing returns only sessions created through the API and excludes sessions created on the page. Single-turn run detail is a GET request carrying appId, chatId and dataId together. The chat interface returns responseData when detail is true, and in non-stream mode the intermediate results are stored there; in stream mode the results are separated by event. Actual token consumption is not returned, so detail must be set to true and the tokens values inside responseData summed manually.
A hypothetical paper exercise sample is set out here for checking only. The material text is: "Inbound plan FBA-2025-0417, planned inbound quantity 600 units, destination warehouse US West warehouse, creation date 17 April 2025, status draft, not yet submitted for platform review." Under the same chatId a second record reads: "Carton label file generated, corresponding to inbound plan FBA-2025-0417, generation date 18 April 2025." The draft sentence to be assessed is: "Inbound plan FBA-2025-0417 passed platform review on 17 April, and the carton label has been synchronized to the freight forwarder."
The checking points follow from the material. The material states the status is draft and not yet submitted for review, so "passed platform review" has no supporting field. The carton label being generated is not the same fact as synchronization to the freight forwarder; a synchronization action would require a separate system implementation, and the material names no such action or receiving party. Both gaps stay open.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and boundary |
|---|---|---|
| Session retrieval scope | Send appId, offset, pageSize in the request body; set source to api | Returns only API-created sessions; page sessions stay outside this list |
| Run detail switch | Set detail=true in the chat request body; read responseData in non-stream mode | Stream mode separates results by event; detail is required for this check |
| Token calculation starting point | Hypothetical exercise starting point: manually sum tokens in responseData and compare with prompt plus completion | The interface does not return actual consumption; the starting point is set for this check only |
| Session identity fields | Check whether chatId and dataId both belong to the target appId | Sessions from different entry points are not merged |
| Interactive node return | Set detail=true when the interactive configuration is needed; read type and params under data.interactive | Internal runtime fields are not returned |
| Plugin output location | Find the element with moduleType=pluginOutput in responseData and read pluginOutput | Plugins run in detail mode by default and do not require chatId |
Three common errors
- Reading a draft status as reviewed. In the hypothetical sample the material text states status draft and not yet submitted for platform review. If the draft sentence states passed platform review, the input is the same inbound plan number and the same status field, the comparison is between the mutually exclusive values draft and passed review, and the observable result is that the draft sentence does not match the material.
- Treating carton label generation as synchronization to the forwarder. The material states only that the carton label file was generated; it names no synchronization action and no receiving party. The draft sentence's "synchronized to the freight forwarder" has no corresponding field in the material and remains unsupported.
- Using a similarity score as a permission or entity filter. In run detail the score reflects recall ordering only. It does not grant authorization over a target business record, and retrieval does not produce an authorization effect.
How to verify the setup
- Take the first material text from the hypothetical sample, request the run detail, and check the query and answer fields for the matching dataId. Compare whether the status description in the record matches draft and not yet submitted for review. If passed review appears, the draft sentence is judged inconsistent with the material.
- Take the second material text from the hypothetical sample and inspect the record for a synchronization action field or a receiving party field. If neither appears after comparison, the phrase "synchronized to the freight forwarder" is judged unsupported and the evidence gap is retained.
- For the session listing request under the same chatId, check whether the returned items' chatId and appId match the target application. If a session from another appId appears after comparison, the retrieval scope is judged incorrect, and the request parameters must be corrected before checking again.