微信扫码
与创始人交个朋友
Atomic Agents是一个不同于CrewAI之类的一个prompt输入+一个按钮点击全自动的Agent框架,面向于真实场景,主打轻量级、灵活、透明的框架
仓库地址如下:
https://github.com/KennyVaneetvelde/atomic_agents
作者自己写的仓库介绍,写的很实在,所以附上翻译直接贴过来了:
I've been working on a new open-source AI agent framework called Atomic Agents. After spending a lot of time on it for my own projects, I became very disappointed with AutoGen and CrewAI.
我一直在研究一个名为 Atomic Agents 的新开源 AI Agents框架。在为自己的项目花费了大量时间之后,我开始对 AutoGen 和 CrewAI 感到非常的失望。
Many libraries try to hide a lot of things and make everything seem magical. They often promote the idea of "Click these 3 buttons and type these prompts, and wow, now you have a fully automated AI news agency." However, these solutions often fail to deliver what you want 95% of the time and can be costly and unreliable.
许多仓库试图隐藏很多东西,让一切看起来都很神奇。他们经常这样宣传:“点击这 3 个按钮并输入这些prompt,哇,现在你有了一个完全自动化的人工智能新闻机构。”然而,这些解决方案通常在 95% 的时间里都无法满足您的需求,而且成本高昂且不可靠。
These libraries try to do too much autonomously, with automatic task delegation, etc. While this is very cool, it is often useless for production. Most production use cases are more straightforward, such as: Search the web for a topic、Get the most promising URLs、Look at those pages、Summarize each page...
这些库试图通过自动任务委派等方式,让整个系统自主完成太多工作。虽然这非常酷,但对于生产来说通常毫无用处。大多数生产用例都更简单,例如:在网络上搜索主题、获取最相关的url、看看那些页面、总结每一页...
To address this, I decided to build my framework on top of Instructor, an already amazing library that constrains LLM output using Pydantic. This allows us to create agents that use tools and outputs completely defined using Pydantic.
为了解决这个问题,我决定在 Instructor 之上构建我的框架,这是一个已经很棒的库,它使用 Pydantic 限制 LLM 输出。这使我们能够创建使用完全由 Pydantic 定义的工具和输出的代理。
Now, to be clear, I still plan to support automatic delegation, in fact I have already started implementing it locally, however I have found that most usecases do not require it and in fact suffer for giving the AI too much to decide.
现在,需要明确的是,我仍然计划支持自动委派,事实上我已经开始在本地实现它,但我发现大多数用例不需要它,而且实际上因为给人工智能太多的决定而受到影响。
The result is a lightweight, flexible, transparent framework that works very well for the use cases I have used it for, even on GPT-3.5-turbo and some bigger local models, whereas autogen and crewAI are complete lost cases unless using only the strongest most expensive models.
最终,Atomic Agents是一个轻量级、灵活、透明的框架,非常适合我使用过的用例,甚至在 GPT-3.5-turbo 和一些更大的本地模型上也是如此,而 autogen 和crewAI则完全失败,除非只使用最强的最昂贵的大模型。
53AI,大模型落地应用首选服务商
定位:开箱即用的大模型落地应用平台
承诺:先做场景POC验证,看到效果再签署服务协议。零风险落地应用大模型,已交付160+中大型企业
如果你看见AI对商业世界的变革,欢迎来和我们一起探索~
2024-07-04
用coze零代码搭建一个知识库助手
2024-07-04
下一代 RAG 技术来了!微软正式开源 GraphRAG:大模型行业将迎来新的升级?
2024-07-04
什么是知识图谱?与大模型有什么关系
2024-07-04
检索生成(RAG) vs 长文本大模型:实际应用中如何选择?
2024-07-04
用腾讯星脉,给AI网络诊脉
2024-07-04
代码 | 使用本地大模型从文本中提取结构化信息
2024-07-04
Karpor 开源了!打造 AI 时代的 Kubernetes 可视化工具
2024-07-04
普林斯顿大学重磅 | 关键的AI Agent,你遇到过这4大Agent评估陷阱吗
2024-04-24
2024-05-14
2024-04-24
2024-04-23
2024-04-11
2024-04-26
2024-03-22
2024-03-29
2024-04-12
2024-04-25
2024-07-04
2024-07-04
2024-07-03
2024-07-02
2024-07-02
2024-07-01
2024-07-01
2024-07-01