Discussion and analysis of Text2SQL technology, the most difficult pain point in the commercial implementation of agents.(Agent商业落地里最难的痛点Text2SQL技术探讨和解析)

Discussion and analysis of Text2SQL technology, the most difficult pain point in the commercial implementation of agents.(Agent商业落地里最难的痛点Text2SQL技术探讨和解析)

Agent商业落地里最难的是Text2SQL(NL2SQL),几乎是无法绕开的核心痛点,主要面临的三个核心问题:

  • 为什么到目前为止仍然没有真正可靠的商业共识性企业级解决方案?
  • 实际企业应用场景中,有哪些靠谱的思路和解决方案?
  • 是依托专有小模型还是基于模版宏套用替换变量的方式?
  • 如果是你,你怎么设计一个准确率足够高的text2sql引擎?

这是一个非常深刻且直击要害的商业落地问题。Text-to-SQL(或者说,更广义的NL2SQL/Text2Analytics)下面我将从“为什么难”、“现有靠谱的思路”以及“技术选型”三个层面,系统地拆解这个问题。

Read more
Ultimate Guide to Quantizing AI Large Language Models: From FP32 to INT4, How to Make Large Models Perform at Full Speed on Consumer Devices?(AI大语言模型量化终极指南:从FP32到INT4,如何让大模型在消费级设备部署应用及选型?)

Ultimate Guide to Quantizing AI Large Language Models: From FP32 to INT4, How to Make Large Models Perform at Full Speed on Consumer Devices?(AI大语言模型量化终极指南:从FP32到INT4,如何让大模型在消费级设备部署应用及选型?)

——深度解析量化格式、尺寸差异与硬件适配策略(附M3 Pro实战指南)

个人常用办公终端设备型号:

  • **Macbook Pro M3 (36G内存定制款) **:

小结

  • 💡 Apple用户闭眼选BF16:M3 Pro芯片的BF16性能碾压FP16,18GB内存可流畅运行30B级模型
  • ⚠️ INT4是双刃剑:70B模型塞进36GB内存的唯一方案,但精度损失高达15%+
  • 🔮 未来属于FP8:NVIDIA H100已支持,苹果M4或成转折点

Read more
Thoughts on Agent-based Enterprise Application Architecture.(Agent企业级应用架构思考和挑战)

Thoughts on Agent-based Enterprise Application Architecture.(Agent企业级应用架构思考和挑战)

“不确定性不是缺陷,而是新范式的特征,必须学会“回忆”,但同时也要学会“遗忘”。”

AI时代,智能体本身的概率输出让软件走向不确定,或者说更个性。但这对企业级产品的准确率形成巨大挑战,怎么看待这种现状、机遇和商业风险?智能体和传统应用范式下在业务落地间角色和职能的划分和原则?

这是目前AI面临的核心问题,触及了AI原生时代企业软件架构、产品设计与组织协作的根本性变革和创业者的产品决策方向。

Read more
AI drawing ControlNet local implementation steps by stable-diffusion-webui(AI绘画ControlNet本地构建实施步骤 by stable-diffusion-webui)

AI drawing ControlNet local implementation steps by stable-diffusion-webui(AI绘画ControlNet本地构建实施步骤 by stable-diffusion-webui)

Install stable-diffusion-webui

[Note]:

  • Unable to install due to network problems, it is recommended to use GIT source pull installation
    (因网络问题导致无法安装,故建议使用GIT源码拉取安装方式)
  • Due to the large size of the model file, it is recommended to ensure sufficient space before installation: At least [source code + data model + extension]25G
    (因模型文件较大,建议安装前确保有足够空间: 至少[源码+数据模型+扩展]25G)
  • The PC terminal to be built is 🍎 macOS Ventura 13.4.1(c), and the Windows environment is built in much the same way.
    (构建PC终端是🍎 macOS Ventura 13.4.1(c),Windows系统环境构建方式大同小异)
  • Basic build tools (brew management recommended)
    (基础构建工具(推荐brew管理))
    1
    2
    3
    4
    5
    6
    cmake
    wget
    git
    rust
    protobuf
    python3.x
Project key directory(项目关键目录)
  • stable-diffusion-webui/extensions (扩展)
  • stable-diffusion-webui/models (模型)

Construction steps(构建步骤)

Read more