Best practices for using AI FDE（使用 AI FDE 的最佳实践）¶

Review the following best practices to help you leverage AI FDE effectively while maintaining system integrity and optimizing results.

Verify resources and procedures¶

Always verify generated resources before implementing them in production environments. This verification process should include reviewing all generated code for correctness and adherence to organizational standards.

Test transformation logic with representative sample data to ensure it performs as expected under various conditions. Note that AI FDE will make changes on a branch by default and propose its changes in a Global Branch proposal or Code Repository pull request for review.

Depending on the tool being used and tool configurations, AI FDE may ask for tool approval before performing an action. Ensure that you understand the action that the model is trying to execute, and that you check the tool being used before approving.

Limit tools and context¶

We recommend that you provide the model with only the context and tools that are essential for a given task. Providing unnecessary context or tools can lead to suboptimal or incorrect actions. By restricting the available tools to only those required for a specific task, and including only the most relevant context, you can reduce the likelihood of confusion or error. Limiting tools and context can improve the efficiency and accuracy of AI FDE, as well as enhance security by minimizing access to sensitive operations or information.

Decompose problems and use iterative development¶

Complex operations should be broken down into smaller, more manageable steps. We recommend that you start with basic structures and gradually add complexity as each component is verified.

AI FDE is particularly effective for rapid prototyping, allowing for quick exploration of different approaches. For production-quality implementations, combine AI-generated foundations with manual development for fine-tuning and optimization.

Track performance with AIP Evals¶

Use AIP Evals to evaluate and track the performance of functions created or modified by AI FDE. Creating evaluation suites for your LLM-backed functions allows you to measure the impact of changes over time and compare different approaches. This is particularly valuable when using AI FDE for iterative function development, as it provides quantitative feedback on whether changes improve or degrade performance.

Consider infrastructure constraints¶

When using AI FDE, it is important to acknowledge the differences in operational patterns compared to human developers. While human developers typically perform one operation at a time, often pausing to think between actions and taking breaks, AI FDE can execute operations in rapid succession, with only seconds between each action. This can trigger dozens of operations in minutes, running continuously until tasks are complete. Multiple AI FDE sessions may operate in parallel, compounding the overall impact on infrastructure. This can expose infrastructure bottlenecks that may be manageable for the slower, more sequential pace of human activity.

Give careful consideration to the following areas:

Storage read/write
Compute resources, particularly GPUs
Storage capacity

AI FDE usage may require infrastructure that can accommodate high-frequency parallel operations, sustained compute loads, increased network activity, and expanded storage needs.

中文翻译¶

使用 AI FDE 的最佳实践¶

请参考以下最佳实践，帮助您在维护系统完整性和优化结果的同时，有效利用 AI FDE。

验证资源与流程¶

在生产环境中实施生成的资源之前，务必先进行验证。此验证过程应包括检查所有生成代码的正确性及其是否符合组织标准。

使用具有代表性的样本数据测试转换逻辑，确保其在各种条件下都能按预期运行。请注意，AI FDE 默认会在分支上进行更改，并通过 Global Branch 提案或 Code Repository 拉取请求提交更改以供审查。

根据所使用的工具及其配置，AI FDE 在执行操作前可能会请求工具审批。请确保您理解模型试图执行的操作，并在批准前检查所使用的工具。

限制工具与上下文¶

我们建议您仅向模型提供特定任务所必需的上下文和工具。提供不必要的上下文或工具可能导致次优或错误的操作。通过将可用工具限制为特定任务所需的工具，并仅包含最相关的上下文，可以降低混淆或出错的可能性。限制工具和上下文不仅能提高 AI FDE 的效率和准确性，还能通过减少对敏感操作或信息的访问来增强安全性。

分解问题并采用迭代开发¶

复杂的操作应分解为更小、更易于管理的步骤。我们建议您从基础结构开始，在验证每个组件后逐步增加复杂性。

AI FDE 在快速原型开发方面尤为有效，可快速探索不同的方法。对于生产级质量的实现，可将 AI 生成的基础与手动开发相结合，进行微调和优化。

使用 AIP Evals 跟踪性能¶

使用 AIP Evals 评估和跟踪由 AI FDE 创建或修改的函数性能。为基于 LLM 的函数创建评估套件，可以衡量变更随时间推移的影响，并比较不同方法。这在使用 AI FDE 进行迭代函数开发时尤其有价值，因为它能提供量化反馈，帮助判断变更是提升还是降低了性能。

考虑基础设施限制¶

使用 AI FDE 时，必须认识到其操作模式与人类开发者的差异。人类开发者通常一次执行一个操作，常常在操作之间暂停思考并休息，而 AI FDE 可以快速连续执行操作，每次操作之间仅间隔数秒。这可能在几分钟内触发数十个操作，并持续运行直至任务完成。多个 AI FDE 会话可能并行运行，从而加剧对基础设施的整体影响。这可能会暴露那些在人类较慢、顺序化的操作节奏下尚可管理的基础设施瓶颈。

请重点关注以下方面：

存储读写
计算资源，尤其是 GPU
存储容量

使用 AI FDE 可能需要能够支持高频并行操作、持续计算负载、增加的网络活动以及扩展存储需求的基础设施。