A Flash Model Ignored Japanese Output—A Three-Step Reliability Fix
The LinkDigest author reports that a low-cost flash model generated Chinese summaries and bullet points for two of the first three Japanese city-report batches, despite an explicit Japanese-output instruction. The proposed fix checks generated prose by character patterns, retries once with a clearer language rule, and falls back to translation while preserving source quotes in code; retries also receive separate token limits and retain prior results on failure. After rerunning eight cities, the author says seven remaining reports had Japanese prose and Chinese quotations, while one report was removed because only one note was found.
The incident and repair pattern provide a concrete case for making multilingual AI output validation, retry limits, deterministic fallbacks, and honest language labeling part of production workflows.
说明:文中的服务是作者自己做的(LinkDigest,把小红书、抖音链接读成文字的接口)。数字来自 2026-11-06 当晚的线上日志。
背景
我们有一个接口:一组小红书笔记读完之后,用一个便宜的 flash 模型写一份「对比报告」(共同主题、开头钩子、读者顾虑、空白点……),报告语言可以指定,比如 translate_to: "ja"。系统提示里写得很清楚:Reply with JSON, in Japanese.
事故
当晚给日本的客户做 8 个城市的公开报告。前 3 批里有 2 批,模型写出来的 summary 和每条要点全是中文。 原因不难猜:输入的笔记是中文,便宜模型被素材的语言「带跑了」,提示词里那句 in Japanese 被无视。
第一反应通常是改提示词:加粗、写三遍、放到最前面。我们的规矩是:一个线上模型两次违反同一条指令,就不再换措辞,改成三步:代码检查 → 一次重问 → 确定性的兜底。