What the documents look like
Amazon FBM fulfillment places fulfillment and return-address configuration with the seller. Platform warehouse inspection does not participate in this flow. The material base for copy drafting therefore comes from order pick lists, shipment confirmation sheets, return address settings, logistics tracking tables, customer service email, and ERP orders. These documents carry in-transit status, return addresses, and tracking numbers. Multi-turn drafting treats those items as the citable scope rather than relying on any platform warehouse conclusion.
Records update per order and per day. Two turns inside one conversation can therefore refer to different order states. When a draft treats a previous turn's shipment confirmation as the current turn's fact, the resulting promotional sentence can contradict the actual order state. This page addresses only whether a proposed copy sentence is supported by the supplied material. Identity certification and report analysis sit outside this task.
Constraints at the multi-turn prompts step
Multi-turn drafting must separate a new conversation from a continued one. When the request body carries a non-empty chatId, the system retrieves the conversation from the database, uses the last element of the messages array as the user question, and ignores the remaining messages. Material text supplied in earlier turns is not referenced unless it entered the history of that chatId. The memory turn setting controls how many earlier turns the model may consult; a small value can prevent earlier information from being used.
The drafting task requires the model to answer within the material scope and to avoid adding unsupported business facts. The reply format follows downstream consumption. Selecting json_schema requires the corresponding format requirement to be filled in and the returned structure to be checked. Retrieval returns content and does not constitute authorization. A similarity score is not a permission filter and cannot be used to exclude unrelated business records or unauthorized records.
Hypothetical drill sample: the source material reads "The pick list for Order A states the item is a cotton short-sleeve shirt, quantity 2; the shipment confirmation sheet states the carrier is a local carrier and the tracking number has been entered into the logistics tracking table; the return address setting is an overseas local address." The draft sentence under review reads "This item supports 30-day no-reason returns, and returns are handled centrally by the platform warehouse." The subject, object, and status needed for the judgment appear in the source text, while "30-day no-reason" and "handled centrally by the platform warehouse" do not appear and remain unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Memory turns | Hypothetical starting value of 4 turns | Drafting needs to look back at source material; compare with 2 turns to observe whether earlier references are lost |
| Reply limit | Hypothetical starting value of 800 | Copy paragraphs run long; compare with 400 to observe whether output is truncated early |
| Temperature | Hypothetical starting value of 0.7 | Copy falls in the creative category; compare with 0.2 to observe variation in wording |
| Reply format | Hypothetical selection of json_schema | Downstream needs to read fixed fields; compare with the default format to observe the returned structure |
| Multimodal recognition | Hypothetical image recognition enabled | Product images need recognition; compare with disabled to observe whether image content enters the input |
| Thinking configuration | Hypothetical selection of standard thinking | The judgment spans several documents; compare with no thinking to observe reasoning completeness |
These values are hypothetical trial starting points for this page and are not official defaults. Where retrieval libraries use different vector models, they cannot be associated together; verify the identifiers first and exclude those that do not match.
Three common errors
- Treating a similarity score as a permission or entity filter. The score reflects content closeness only and cannot decide whether a record belongs to the current business or has been authorized. Retrieved results still require identifier checks.
- Treating a pending review as an approval. Submitting a draft sentence for review only means it entered the review stage. It does not mean the copy was approved or published, and it does not mean a return policy has taken effect.
- Treating related-stage evidence in the material as a contradiction in the same field. When the material does not state a return period, the correct record is unsupported or pending verification, not a negative finding or a conflict.
How to verify the setup
- Input the source material and the draft sentence from the hypothetical sample, then compare whether "30-day no-reason returns" and "handled centrally by the platform warehouse" appear in the material. Neither appears, so the expected result is unsupported, and the evidence gap is recorded.
- Input a user question containing an image link, then compare whether the model input includes image content before and after image recognition is enabled. Only when the model itself supports images and the link falls within 500 characters and no more than 4 media links are present is the expected result that image content enters the input.
- Input the paper input [] for the empty-result handoff, then compare whether the downstream node follows the empty-reference design. The expected result is that the downstream node produces no reference content. Do not claim that a fictional rule retrieval necessarily returns [].