Show HN: 我搭建了一个基于本地优先的、具备高级 RAG 功能的 AI 研究平台
4 分•作者: vinhnx•6 个月前
我开发了 VT Chat,这是一个注重隐私的 AI 平台,所有对话都在本地进行,同时提供高级研究功能,并可访问 15 种以上的 AI 模型,包括 Claude 4 Sonnet 和 Claude 4 Opus、O3、Gemini 2.5 Pro 和 DeepSeek R1。
在线体验:<a href="https://vtchat.io.vn" rel="nofollow">https://vtchat.io.vn</a> | 源代码:<a href="https://github.com/vinhnx/vtchat">https://github.com/vinhnx/vtchat</a>
其核心区别在于真正的本地优先架构。所有聊天记录都存储在 IndexedDB 中,零服务器存储。您的 API 密钥永远不会离开您的浏览器,即使作为开发者,我也无法看到它们。您可以在 OpenAI、Anthropic、Google 和其他提供商之间切换,以比较回复,并在共享机器上实现完全的数据隔离。
我添加了自己每天都在使用的研究功能:深度研究 (Deep Research) 通过来源验证进行多步骤研究,专业搜索 (Pro Search) 集成实时网络搜索,AI 记忆 (AI Memory) 从您的对话中创建个人知识库。此外,还有用于 PDF 的文档处理功能,一个用于查看完整 AI 推理的“思考模式”,以及将文档转换为 JSON 的结构化提取功能。AI 驱动的语义路由会根据您的查询自动激活工具。
使用 Next.js 14、TypeScript 和 Turborepo 在单体仓库中构建。所有内容都完全开源,可使用您自己的 API 密钥进行自托管。托管版本保留了大多数免费功能,并提供可选的 VT+,用于高级模型和高级研究功能。
欢迎讨论本地优先设计决策、研究工作流程架构或任何其他技术细节。
查看原文
I've built VT Chat, a privacy-first AI platform that keeps all your conversations local while providing advanced research capabilities and access to 15+ AI models including Claude 4 Sonnet and Claude 4 Opus, O3, Gemini 2.5 Pro and DeepSeek R1.<p>Live: <a href="https://vtchat.io.vn" rel="nofollow">https://vtchat.io.vn</a> | Source: <a href="https://github.com/vinhnx/vtchat">https://github.com/vinhnx/vtchat</a><p>The core difference is true local-first architecture. All chats are stored in IndexedDB with zero server storage. Your API keys never leave your browser, and I can't see them even as the developer. You can switch between OpenAI, Anthropic, Google and other providers to compare responses, with complete data isolation for shared machines.<p>I've added research features that I actually use daily: Deep Research does multi-step research with source verification, Pro Search integrates real-time web search, and AI Memory creates a personal knowledge base from your conversations. There's also document processing for PDFs, a "thinking mode" to see complete AI reasoning, and structured extraction to turn documents into JSON. AI-powered semantic routing automatically activates tools based on your queries.<p>Built with Next.js 14, TypeScript, and Turborepo in a monorepo setup. Everything is fully open source for self-hosting with your own API keys. The hosted version keeps most features free, with optional VT+ for premium models and advanced research capabilities.<p>Happy to discuss the local-first design decisions, research workflow architecture, or any other technical details.