Prompt 模板化
生产环境不建议让每个业务方随意写 Prompt。更可靠的方法是把 Prompt 拆成角色、任务、输入、输出格式、限制条件和错误处理。
缓存能降低重复成本
很多企业任务存在重复上下文，例如固定产品资料、政策文件、FAQ 或系统说明。合理利用缓存输入价格，可以降低长上下文任务的重复成本。
路由策略
不是所有任务都需要最高推理强度。简单分类、格式转换、短文本润色可以使用较低配置；复杂分析、代码工程、跨文档推理再使用更高推理强度。
- 低复杂度：格式整理、标签分类、简单摘要。
- 中复杂度：知识库问答、邮件草稿、客服建议。
- 高复杂度：多文档分析、代码工程、业务 Agent 规划。
评测与监控
每个高频场景都应该保留一组样本问题，定期评估回答质量、格式稳定性、延迟和 token 消耗。
Template prompts
Production systems should not let every team write free-form prompts. A more reliable pattern separates role, task, input, output format, constraints and error handling.
Caching reduces repeated cost
Many enterprise workflows repeat context, such as product data, policies, FAQs or system instructions. Using cached input effectively can lower repeated long-context cost.
Routing strategy
Not every task needs maximum reasoning. Simple classification, formatting and short rewriting can use lower settings; complex analysis, code work and cross-document reasoning can use higher reasoning effort.
- Low complexity: formatting, tagging and short summaries.
- Medium complexity: knowledge Q&A, email drafts and support suggestions.
- High complexity: multi-document analysis, code engineering and agent planning.
Evaluation and monitoring
Every high-frequency scenario should keep sample questions and regularly measure quality, format stability, latency and token usage.
需要优化现有模型 API 成本和稳定性？Need to optimize API cost and reliability?
提交优化需求Submit an optimization request