After-sales repairs and reships: multi-turn prompts for copy drafting

Review multi-turn prompts for copy drafting in After-sales repairs and reships. Check input mappings, source evidence and failure handling before use.

What the documents look like

After-sales repair and reship work aims to restore a product to use, and it closes on a completed repair or a delivered replacement part. The main materials are the repair request, the reship order, the inspection report and the parts warehouse outbound order. They update per order and per day. Copy drafting draws the subject, the product object, the fault state, the handling method and the valid scope from these documents, then turns them into customer-facing statements or marketing wording. For the multi-turn prompts step, the history carried into each turn, the knowledge references and the model output all need to be checkable, so an unconfirmed repair conclusion never reads as a completed fact.

Constraints at the multi-turn prompts step

Multi-turn dialogue separates a new session from a continued session. When chatId is not passed, the platform context is not used and the context is built entirely from messages. When a non-empty chatId is passed, the platform retrieves the session automatically and uses only the last item of the messages array as the user question. The prompt states the material scope, the answering task and how to behave when evidence is missing. The amount of history carried in is governed by the memory turns setting; a larger value adds consumption and affects response speed. The reply format follows what the downstream node needs to consume; choosing json_schema requires filling in the corresponding format requirements and checking the returned structure. A relevance score from retrieval does not carry permission or entity filtering duties, and retrieval itself is not authorization.

Hypothetical paper exercise sample: the material text reads "Repair request WX-2024-0517, subject customer A, product a certain model of handheld vacuum, fault described as failing to start after charging, after-sales outlet initial inspection date 17 May 2024, current status awaiting inspection, inspection report not yet issued." A proposed draft sentence reads "The vacuum has been confirmed to have a damaged motor and will be repaired and shipped back within three days." Checking shows the material records only awaiting inspection and no issued report, so the damaged-motor conclusion and the three-day deadline in the draft sentence have no support. The sentence should instead state that the item is at the awaiting-inspection stage pending the inspection report.

Configuration and implementation checks

The table lists only settings the official documents describe, or implementation checks, that relate to this page's copy drafting task. The hypothetical trial starting point is for this paper comparison only and is not an official default.

Configuration itemHypothetical trial starting pointReason and comparison method
AI modelChoose a model supporting multimodal recognitionInspection reports may contain images; compare recognition results against a model without multimodal support
Memory turnsSet to 4 turnsCustomer-service applications usually keep a small amount of history; compare context completeness against 2 turns and 8 turns
TemperatureSet lowCopy needs to stay aligned with the material; compare wording divergence against a higher value
Reply formatKeep defaultNo downstream fixed-structure read is needed; compare downstream handling convenience against json_schema
Multimodal recognitionEnable image recognitionImages in the inspection report need to be read; compare information completeness against the disabled state
Hide AI thinkingEnableThe customer-facing output needs only the final copy; compare intermediate content exposure against the disabled state

Three common errors

  1. Writing awaiting inspection as confirmed damage. The input is the material text "current status awaiting inspection, inspection report not yet issued." Compare the proposed draft sentence "confirmed to have a damaged motor." The observable result is that the draft sentence's conclusion goes beyond what the material records.
  2. Treating an unissued inspection report as support for a repair deadline. The input is the absence of an inspection report in the material. Compare the proposed draft sentence "repaired and shipped back within three days." The observable result is that the deadline lacks material support.
  3. Treating a retrieval relevance score as a permission filter. The input is a similarity score returned by one knowledge-base retrieval. Compare whether unrelated subject records are excluded on that basis. The observable result is that the score carries no entity or permission isolation duty, and identifiers need separate checking.

How to verify the setup

  1. For the hypothetical sample material text and draft sentence, input the "awaiting inspection" and "inspection report not yet issued" fields from the material. Compare whether the draft sentence contains "confirmed damage." If it does, the sentence is judged to exceed material support.
  2. For the "within three days" phrase in the hypothetical sample draft sentence, input the material's inspection report status of not yet issued. Compare whether a repair deadline basis exists. If none exists, the phrase is judged unsupported and the evidence gap is retained.
  3. For empty-result handling, input the paper retrieval result []. Compare whether the downstream node outputs according to the no-reference design. Observe whether the downstream directly fabricates a repair conclusion. If it does, it is judged as not handling the empty result as designed.