RΞASON框架适合构建LLM应用吗？它的TypeScript结构化输出好用吗？

引言：LLM 作为一种新的编程原语

I view Large Language Models (LLMs) as a transformative programming primitive, comparable to the advent of databases in the 1970s or the explosion of web technologies in the early 2000s. They grant developers a unique capability: the power to infuse applications with reasoning and natural language understanding directly within code.

我将大型语言模型（LLMs）视为一种变革性的编程原语，其重要性堪比 20 世纪 70 年代的数据库或 21 世纪初的 Web 技术爆炸。它们为开发者提供了一种独特的能力：将推理和自然语言理解直接融入代码的能力。

However, new primitives demand new abstractions. The goal of RΞASON is to provide precisely that—a robust, intuitive abstraction layer that seamlessly integrates LLMs into the modern developer's workflow.

然而，新的原语需要新的抽象。RΞASON 的目标正是提供这样一个强大、直观的抽象层，将 LLM 无缝集成到现代开发者的工作流中。

什么是 RΞASON？

RΞASON is an open-source TypeScript backend framework designed specifically for building production-grade LLM applications. Its core philosophy is to leverage familiar programming concepts rather than forcing developers to learn entirely new paradigms.

RΞASON 是一个开源的 TypeScript 后端框架，专为构建生产级 LLM 应用而设计。其核心理念是利用熟悉的编程概念，而不是强迫开发者学习全新的范式。

One of its most distinctive features is its use of TypeScript interface definitions to obtain structured, type-safe output from LLMs, eliminating the need for complex and brittle text parsing.

其最显著的特点之一是使用 TypeScript 的 interface 定义来从 LLM 获取结构化、类型安全的输出，从而消除了对复杂且脆弱的文本解析的需求。

快速开始

To get started with RΞASON, install it using the following command:

要开始使用 RΞASON，请使用以下命令安装：

npx use-reason@latest

Next, create a simple API endpoint. For example, create a file at src/entrypoints/joke.ts:

接下来，创建一个简单的 API 端点。例如，在 src/entrypoints/joke.ts 创建一个文件：

import { reasonStream } from 'tryreason'

interface Joke {
  joke: string;
  topics: string[]
}

export async function* GET() {
  return reasonStream<Joke>('tell me a joke')
}

Finally, run the development server:

最后，运行开发服务器：

npm run dev

This blog post will delve into the core features and design philosophy of RΞASON. For a comprehensive tutorial, please refer to the official RΞASON documentation.

本文将深入探讨 RΞASON 的核心功能和设计理念。如需完整的教程，请参阅官方的 RΞASON 文档。

设计背景与核心理念

While LLMs represent a new primitive, they should not necessitate a complete overhaul of existing programming paradigms. The key is to adapt them into our current workflows. RΞASON is built upon five essential principles to achieve this integration.

虽然 LLM 代表了一种新的原语，但它们不应要求彻底改变现有的编程范式。关键在于将它们适配到我们当前的工作流中。RΞASON 基于五个基本原则来实现这种集成。

结构化输出通过特定格式或模式控制LLM输出结构的技术，通常与思维链推理结合使用，提高答案的准确性和一致性。至关重要，但不应以牺牲开发体验为代价

The fundamental input and output of LLMs is unstructured text. While ideal for human interaction, this poses significant challenges for programmatic use:

LLM 的基本输入和输出是非结构化的文本。虽然这对人机交互很理想，但对程序化使用却带来了重大挑战：

• 字符串解析困难 (String parsing is hard): Extracting structured data from natural language requires complex, error-prone parsing logic.
• 难以维护 (Hard-to-maintain): Parsing logic must be frequently updated to handle the variability and nuance of LLM outputs.

• 字符串解析困难: 从自然语言中提取结构化数据需要复杂且容易出错的解析逻辑。
• 难以维护: 解析逻辑必须频繁更新，以处理 LLM 输出的多变性和细微差别。

Many frameworks address structured output but introduce new, proprietary APIs. RΞASON takes a different approach by utilizing the familiar TypeScript interface and JSDoc comments.

许多框架解决了结构化输出通过特定格式或模式控制LLM输出结构的技术，通常与思维链推理结合使用，提高答案的准确性和一致性。的问题，但引入了新的、专有的 API。RΞASON 采用了一种不同的方法，利用熟悉的 TypeScript interface 和 JSDoc 注释。

import { reason } from 'tryreason'

interface Joke {
  /** Use this property to indicate the age rating of the joke */
  rating: number;
  joke: string;
  /** Use this property to explain the joke to those who did not understand it */
  explanation: string;
}

const joke = await reason<Joke>('tell me a really spicy joke')

The reason() function directly queries an LLM and returns a fully typed object. For instance, the joke variable above would yield:

reason() 函数直接查询 LLM 并返回一个完全类型化的对象。例如，上面的 joke 变量将产生如下输出：

{
  "joke": "I'd tell you a chemistry joke but I know I wouldn't get a reaction.",
  "rating": 18,
  "explanation": "This joke is a play on words. The term 'reaction' refers to both a chemical process and a response from someone. The humor comes from the double meaning, implying that the joke might not be funny enough to elicit a response."
}

提示工程应由开发者主导

A framework should assist in areas that are not core differentiators for your application. In the LLM space, many frameworks offer pre-built prompts, agents, and retrieval strategies. However, these elements are often critical to an app's success and should be under the developer's control, not the framework's.

一个框架应该在那些并非你应用核心差异化的领域提供帮助。在 LLM 领域，许多框架提供了预构建的提示词、智能体和检索策略。然而，这些元素通常对应用的成功至关重要，应该由开发者控制，而不是框架。

Furthermore, given the rapid evolution of LLM techniques, frameworks that bundle pre-made solutions risk either becoming outdated or bloated as they struggle to support both old and new methodologies.

此外，鉴于 LLM 技术的快速发展，捆绑预制解决方案的框架要么有过时的风险，要么会变得臃肿，因为它们需要同时支持新旧方法。

Consequently, RΞASON does not provide pre-packaged prompts or agent templates. Instead, it focuses on providing excellent tools and patterns for developers to craft their own.

因此，RΞASON 不提供预打包的提示词或智能体模板。相反，它专注于为开发者提供优秀的工具和模式，让他们能够构建自己的解决方案。

智能体应是函数式的，并能与代码互操作

Many frameworks offer agent creation through class-based templates like new ConversationalAgent(). This approach has the pitfalls mentioned earlier: potential obsolescence and API lock-in.

许多框架通过基于类的模板（如 new ConversationalAgent()）来提供智能体创建功能。这种方法存在前面提到的缺陷：潜在的过时性和 API 锁定。

The alternative—custom agent creation—is more flexible but often leads to excessive boilerplate code in Object-Oriented Programming (OOP) approaches, as seen in libraries like LangChain.

另一种选择——自定义智能体创建——更加灵活，但在面向对象编程（OOP）方法中通常会导致过多的样板代码，正如在 LangChain 等库中所见。

RΞASON addresses this by modeling agents and actions as standard JavaScript/TypeScript functions. This functional approach reduces boilerplate and enhances interoperability. Here's how a web search agent looks in RΞASON:

RΞASON 通过将智能体和动作建模为标准 JavaScript/TypeScript 函数来解决这个问题。这种函数式方法减少了样板代码并增强了互操作性。以下是一个网络搜索智能体在 RΞASON 中的示例：

import { useAgent } from 'tryreason'
import serpApi from '../actions/serp-api'

export const actions = [ serpApi ]

/**
 * You are a helpful assistant that can answer questions about current events.
 */
export default async function WebAgent(userMessage: string) {
  const agent = await useAgent()
  return agent.run(userMessage)
}

When the LLM selects an action, RΞASON simply invokes the corresponding JavaScript function with the LLM-generated parameters, waits for the result, and passes it back to the LLM. Both agents and actions are just functions, leading to a cleaner and more composable architecture.

当 LLM 选择一个动作时，RΞASON 只需使用 LLM 生成的参数调用相应的 JavaScript 函数，等待结果，然后将其传回给 LLM。智能体和动作都只是函数，这带来了更清晰、更可组合的架构。

流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。：不仅仅是文本

LLM completions can be computationally intensive, often taking many seconds. Streaming responses is therefore essential for a good user experience. However, streaming plain text is insufficient; developers need to stream structured data for the reasons outlined in the "Structured Output" section.

LLM 生成可能是计算密集型的，通常需要数秒时间。因此，流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。响应对于良好的用户体验至关重要。然而，仅流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。纯文本是不够的；出于“结构化输出通过特定格式或模式控制LLM输出结构的技术，通常与思维链推理结合使用，提高答案的准确性和一致性。”部分所述的原因，开发者需要流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。结构化数据。

RΞASON supports streaming structured outputs natively via the reasonStream function.

RΞASON 通过 reasonStream 函数原生支持流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。结构化输出通过特定格式或模式控制LLM输出结构的技术，通常与思维链推理结合使用，提高答案的准确性和一致性。。

import { reasonStream } from 'tryreason';

interface City {
  description: string;
  state: string;
  country: string;
  population: number;
}

export async function* GET(req: Request) {
  return reasonStream<City>('tell me about San Francisco')
}

开箱即用的可观测性

To build a successful LLM application, observability is paramount. Developers need insights into user queries, LLM responses, latency, agent execution paths, action calls, and more.

要构建一个成功的 LLM 应用，可观测性至关重要。开发者需要深入了解用户查询、LLM 响应、延迟、智能体执行路径、动作调用等信息。

RΞASON provides zero-setup observability by being natively compatible with OpenTelemetry一个开源的观测性框架，用于收集、处理和导出遥测数据（如指标、日志和追踪），RΞASON内置兼容此框架以实现可观测性。. No additional packages, decorators, or configuration are required. Simply using RΞASON grants you detailed tracing and monitoring capabilities out of the box.

RΞASON 通过原生兼容 OpenTelemetry一个开源的观测性框架，用于收集、处理和导出遥测数据（如指标、日志和追踪），RΞASON内置兼容此框架以实现可观测性。 提供了零配置可观测性。无需额外的包、装饰器或配置。只需使用 RΞASON，您就能获得开箱即用的详细追踪和监控能力。

框架对比与总结

To clearly illustrate RΞASON's positioning against common patterns in other frameworks, the following table highlights its key differentiators:

为了清晰地说明 RΞASON 相对于其他框架常见模式的定位，下表突出了其主要差异化特点：

结论

RΞASON is still in its early stages, and as with any young project, encountering bugs is possible. The development team is committed to addressing issues promptly. Feedback and contributions are highly encouraged—if you encounter any problems, please file an issue on GitHub.

RΞASON 仍处于早期阶段，与任何年轻项目一样，可能会遇到错误。开发团队致力于迅速解决问题。我们非常鼓励反馈和贡献——如果您遇到任何问题，请在 GitHub 上提交 issue。

The vision for RΞASON is to become the foundational layer upon which developers can build sophisticated, reliable, and maintainable LLM-powered applications using the tools and paradigms they already know and love. We are excited to see what the community will build with it.

RΞASON 的愿景是成为一个基础层，让开发者能够使用他们已知且喜爱的工具和范式，构建复杂、可靠且可维护的由 LLM 驱动的应用程序。我们期待看到社区用它构建出怎样的作品。

常见问题（FAQ）

RΞASON框架一个开源的TypeScript后端框架，专门用于构建LLM应用程序，强调结构化输出、函数式代理和开发者对核心逻辑的完全控制。如何实现结构化输出通过特定格式或模式控制LLM输出结构的技术，通常与思维链推理结合使用，提高答案的准确性和一致性。？

RΞASON通过TypeScript接口TypeScript中的一种类型定义方式，在RΞASON中用于指定LLM输出的数据结构，确保类型安全和结构化。定义直接从LLM获取结构化、类型安全的输出，无需复杂的文本解析，开发者只需定义interface即可获得类型化对象。

RΞASON的智能体设计有什么特点？

RΞASON采用函数式智能体设计，能与代码互操作，强调开发者主导的提示工程，将LLM作为编程原语集成到现有工作流中。

RΞASON如何支持流式传输一种数据传输方式，允许数据在生成时逐步发送，而不是等待全部完成，RΞASON支持流式传输结构化数据以优化LLM响应体验。？

RΞASON提供原生流式支持，通过reasonStream函数实现，不仅支持文本流，还能处理结构化数据的流式输出，提升应用响应性。