Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Infinite Research / 无限研究

A recursive BFS-powered research engine that leverages Vane for AI-driven web search, automatically generating and exploring sub-questions to build a comprehensive knowledge tree.
基于递归广度优先搜索的研究引擎，利用 Vane 进行 AI 驱动的网络搜索，自动生成并探索子问题，构建完整的知识树。

Table of Contents / 目录

Features / 功能特性
Architecture / 架构
Message Processing Flow / 消息处理流程
Prerequisites / 前置要求
Vane Installation / Vane 安装
App Installation / 应用安装
Configuration / 配置
Usage / 使用方法
Project Structure / 项目结构

Features / 功能特性

Feature	功能
Recursive BFS search with auto-generated sub-questions	递归广度优先搜索，自动生成子问题
Interactive ReactFlow canvas with task tree visualization	基于 ReactFlow 的交互式任务树可视化画布
TF-IDF cosine similarity deduplication (threshold 0.7)	TF-IDF 余弦相似度去重（阈值 0.7）
Up to 3 concurrent search tasks	最多 3 个并发搜索任务
Sources sidebar with flat/grouped views and content snippets	来源侧边栏，支持平铺/分组视图和内容摘要
Execution logs with level filtering	执行日志，支持级别过滤
Session history with IndexedDB persistence	会话历史，IndexedDB 持久化存储
Export to Markdown, JSON, and TSV formats	支持导出 Markdown、JSON、TSV 格式
Pause / Resume / Stop controls	暂停 / 继续 / 停止控制

Architecture / 架构

graph TB
subgraph Browser ["Browser / 浏览器"]
UI["InputBar<br/>用户输入"]
Canvas["ReactFlow Canvas<br/>任务树画布"]
Sources["SourcesSidebar<br/>来源侧边栏"]
History["HistorySidebar<br/>历史侧边栏"]
Logs["LogsPanel<br/>日志面板"]
StatusBar["StatusBar<br/>状态栏"]
subgraph Engine ["TaskEngine / 任务引擎"]
Queue["BFS Queue<br/>广度优先队列"]
Runner["Concurrent Runner<br/>并发执行器 (max 3)"]
Dedup["TF-IDF Dedup<br/>TF-IDF 去重"]
end
subgraph Storage ["Storage / 存储"]
IDB["IndexedDB<br/>sessions + embeddings"]
end
end
subgraph Backend ["Vane Backend / Vane 后端"]
VaneAPI["Vane API<br/>/api/search"]
SearxNG["SearxNG<br/>搜索聚合"]
LLM["Local LLM<br/>本地大模型"]
end
UI -->|"query / 查询"| Engine
Engine -->|"tasks, logs / 任务, 日志"| Canvas
Engine -->|"tasks, logs"| Sources
Engine -->|"tasks, logs"| Logs
Engine -->|"stats / 统计"| StatusBar
Engine -->|"save / 保存"| IDB
History -->|"load session / 加载会话"| IDB
Runner -->|"POST /api/search"| VaneAPI
VaneAPI --> SearxNG
VaneAPI --> LLM
style Engine fill:#1e3a5f,stroke:#3b82f6
style Storage fill:#1e3a2f,stroke:#10b981
style Backend fill:#3a1e3f,stroke:#a855f7

Component Overview / 组件概览

src/
├── App.jsx # Main app, ReactFlow canvas, state management
│ # 主应用，ReactFlow 画布，状态管理
├── lib/
│ ├── searchApi.js # Vane API client, source normalization
│ │ # Vane API 客户端，来源数据标准化
│ ├── taskEngine.js # BFS engine, concurrency, logging, persistence
│ │ # BFS 引擎，并发控制，日志记录，持久化
│ ├── similarity.js # TF-IDF cosine similarity deduplication
│ │ # TF-IDF 余弦相似度去重
│ └── storage.js # IndexedDB wrapper (sessions + embeddings)
│ # IndexedDB 封装（会话 + 嵌入向量存储）
├── components/
│ ├── InputBar.jsx # Search input, controls, history toggle
│ │ # 搜索输入，控制按钮，历史切换
│ ├── TaskNode.jsx # ReactFlow custom node for each task
│ │ # ReactFlow 自定义任务节点
│ ├── SourcesSidebar.jsx # Right sidebar: sources + logs tabs
│ │ # 右侧边栏：来源 + 日志标签页
│ ├── HistorySidebar.jsx # Left sidebar: session history
│ │ # 左侧边栏：会话历史
│ ├── LogsPanel.jsx # Filterable execution log viewer
│ │ # 可过滤的执行日志查看器
│ └── StatusBar.jsx # Bottom stats bar
│ # 底部状态栏
└── styles/
└── index.css # Tailwind directives + animations
# Tailwind 指令 + 动画

Message Processing Flow / 消息处理流程

sequenceDiagram
participant U as User / 用户
participant E as TaskEngine / 任务引擎
participant V as Vane API
participant D as TF-IDF Dedup / 去重
participant S as IndexedDB
U->>E: Submit query / 提交查询
E->>E: Create root task (depth 0)<br/>创建根任务（深度 0）
E->>E: Add to BFS queue / 加入队列
loop BFS Processing / BFS 处理 (max 3 concurrent / 最多3并发)
E->>V: POST /api/search (query)
V-->>E: {message, sources[]}
E->>E: Store result + sources<br/>保存结果和来源
E->>V: POST /api/search (generate sub-tasks prompt)
V-->>E: JSON array of 1-3 sub-questions<br/>返回1-3个子问题
loop For each sub-question / 对每个子问题
E->>D: isDuplicate(subQuery, existingTasks, 0.7)
alt Similarity > 0.7 / 相似度 > 0.7
D-->>E: duplicate=true
E->>E: Mark as skipped / 标记为跳过
else Similarity <= 0.7 / 相似度 <= 0.7
D-->>E: duplicate=false
E->>E: Create task, add to queue<br/>创建任务，加入队列
end
end
E->>S: Debounced save (500ms)<br/>防抖保存（500毫秒）
E-->>U: Update UI (tasks, logs, stats)<br/>更新界面（任务，日志，统计）
end

Task Lifecycle / 任务生命周期

stateDiagram-v2
[*] --> pending: Created / 创建
pending --> running: Dequeued / 出队执行
running --> completed: Search + sub-tasks done / 搜索+子任务完成
running --> failed: Error / 错误
pending --> skipped: Duplicate detected / 检测到重复
completed --> [*]
failed --> [*]
skipped --> [*]

Prerequisites / 前置要求

Requirement / 要求	Version / 版本
Node.js	>= 18
npm	>= 9
Vane (running instance)	latest
Ollama (for Vane's LLM backend)	latest
A chat model in Ollama	e.g. `llama3`, `gemma2`, etc.
An embedding model in Ollama	e.g. `nomic-embed-text`, etc.

Vane Installation / Vane 安装

Vane is a privacy-focused AI search engine that runs on your own hardware. It provides the search API that powers Infinite Research.

Vane 是一个注重隐私的 AI 搜索引擎，完全在本地硬件上运行。它为 Infinite Research 提供搜索 API。

Full documentation / 完整文档: github.com/ItzCrazyKns/Vane

Option A: Docker (Recommended / 推荐)

# Pull and run with bundled SearxNG# 拉取并运行（内置 SearxNG）
docker run -d \
-p 3000:3000 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:latest

For custom SearxNG instance / 使用自定义 SearxNG 实例:

docker run -d \
-p 3000:3000 \
-e SEARXNG_API_URL=http://your-searxng:8080 \
-v vane-data:/home/vane/data \
--name vane \
itzcrazykns1337/vane:slim-latest

Option B: From Source / 从源码安装

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane
npm i
npm run build
npm run start

Configure Ollama for Vane / 为 Vane 配置 Ollama

Install Ollama / 安装 Ollama: ollama.com

Pull required models / 拉取所需模型:

ollama pull <your-chat-model># e.g. llama3, gemma2, etc. / 对话模型
ollama pull <your-embedding-model># e.g. nomic-embed-text, etc. / 嵌入模型

Ensure Ollama listens on 0.0.0.0 (required for Docker or remote access): 确保 Ollama 监听 0.0.0.0（Docker 或远程访问时必需）:

Linux:

# Edit /etc/systemd/system/ollama.service, add:# 编辑 /etc/systemd/system/ollama.service，添加：
Environment="OLLAMA_HOST=0.0.0.0:11434"
systemctl daemon-reload
systemctl restart ollama

macOS / Windows:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Open Vane at http://localhost:3000 and configure: 打开 Vane (http://localhost:3000) 并配置：
- Select your Ollama chat model / 选择 Ollama 对话模型
- Select your Ollama embedding model / 选择嵌入模型
- Note the Provider ID shown in Vane's settings / 记下设置中显示的 Provider ID

App Installation / 应用安装

# Clone the repository / 克隆仓库
git clone <repository-url>cd gemma_claude
# Install dependencies / 安装依赖
npm install
# Copy env template and edit / 复制环境变量模板并编辑
cp .env.example .env
# Start development server / 启动开发服务器
npm run dev

The app will be available at http://localhost:5180.

应用将在 http://localhost:5180 启动。

Configuration / 配置

All configuration is managed through the .env file. Copy .env.example to .env and edit the values.

所有配置通过 .env 文件管理。将 .env.example 复制为 .env 并编辑其中的值。

cp .env.example .env

Environment Variables / 环境变量

Variable / 变量	Default / 默认值	Description / 说明
`VITE_API_BASE`	`/api`	API base path (usually no need to change) / API 基础路径（通常无需修改）
`VITE_VANE_TARGET`	`http://localhost:3000`	Vane instance URL / Vane 实例地址
`VITE_CHAT_PROVIDER_ID`	—	Vane chat provider UUID / Vane 对话提供者 UUID
`VITE_CHAT_MODEL_KEY`	—	Chat model name in Ollama / Ollama 中的对话模型名称
`VITE_EMBED_PROVIDER_ID`	—	Vane embedding provider UUID / Vane 嵌入提供者 UUID
`VITE_EMBED_MODEL_KEY`	—	Embedding model name / 嵌入模型名称
`VITE_MAX_CONCURRENT`	`3`	Max parallel search tasks / 最大并行搜索任务数
`VITE_SIMILARITY_THRESHOLD`	`0.7`	TF-IDF dedup threshold (0–1) / TF-IDF 去重阈值
`VITE_MAX_SUB_TASKS`	`3`	Max sub-questions per task / 每个任务最大子问题数
`VITE_FETCH_TIMEOUT`	`600000`	API request timeout in ms (10 min) / API 请求超时（毫秒）
`VITE_PORT`	`5180`	Dev server port / 开发服务器端口

Example `.env` / 示例

# Vane APIVITE_API_BASE=/apiVITE_VANE_TARGET=http://localhost:3000# Models (find Provider ID in Vane settings)# 模型配置（在 Vane 设置页面找到 Provider ID）VITE_CHAT_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_CHAT_MODEL_KEY=your-chat-modelVITE_EMBED_PROVIDER_ID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxxVITE_EMBED_MODEL_KEY=your-embedding-model# Engine tuning / 引擎调优VITE_MAX_CONCURRENT=3VITE_SIMILARITY_THRESHOLD=0.7VITE_MAX_SUB_TASKS=3VITE_FETCH_TIMEOUT=600000# Dev server / 开发服务器VITE_PORT=5180

Finding your Provider ID / 查找 Provider ID: Open Vane's web UI, go to settings. The provider ID is a UUID shown next to each configured provider (Ollama, OpenAI, etc.).
打开 Vane 网页界面，进入设置页面。Provider ID 是每个配置的提供者（Ollama、OpenAI 等）旁边显示的 UUID。

Usage / 使用方法

Start a research / 开始研究: Enter a research question in the top search bar and click "Research". 在顶部搜索栏输入研究问题，点击 "Research"。
Watch the tree grow / 观察树的生长: The ReactFlow canvas shows tasks as nodes. Running tasks pulse, completed tasks turn green, skipped duplicates turn yellow. ReactFlow 画布以节点形式展示任务。运行中的任务闪烁，已完成的变绿，跳过的重复项变黄。
Control execution / 控制执行:
- Pause / 暂停 - Stops dequeuing new tasks (running tasks finish) / 停止出队新任务（运行中的任务会完成）
- Resume / 继续 - Resumes queue processing / 恢复队列处理
- Stop / 停止 - Stops all processing / 停止所有处理
Browse sources / 浏览来源: Click "Sources" to open the right sidebar. Toggle between flat and grouped views. Filter by keyword. Expand cards to see full content snippets. 点击 "Sources" 打开右侧边栏。切换平铺和分组视图。按关键词过滤。展开卡片查看完整内容摘要。
View logs / 查看日志: Switch to the "Logs" tab in the right sidebar. Filter by level (Info/Warn/Error). 在右侧边栏切换到 "Logs" 标签页。按级别过滤（Info/Warn/Error）。
Export results / 导出结果:
- Markdown (.md) - Full research report / 完整研究报告
- JSON (.json) - Structured task data / 结构化任务数据
- TSV (.tsv) - Source links spreadsheet / 来源链接表格
Session history / 会话历史: Click "History" to browse past sessions. Click any session to reload it on the canvas. Sessions persist across page refreshes via IndexedDB. 点击 "History" 浏览历史会话。点击任意会话在画布上重新加载。会话通过 IndexedDB 在页面刷新后持久保存。

Tech Stack / 技术栈

Layer / 层	Technology / 技术
UI Framework	React 19
Build Tool	Vite 6
Styling	TailwindCSS 3
Graph Visualization	@xyflow/react (ReactFlow) 12
Search Backend	Vane (SearxNG + LLM)
LLM	Ollama (any compatible model)
Persistence	IndexedDB (browser-native)
Deduplication	TF-IDF Cosine Similarity

License / 许可证

MIT

Name		Name	Last commit message	Last commit date
.claude		.claude
src		src
.DS_Store		.DS_Store
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.js		tailwind.config.js
vite.config.js		vite.config.js

Name		Name	Last commit message	Last commit date
.claude		.claude
src		src
.DS_Store		.DS_Store
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.js		tailwind.config.js
vite.config.js		vite.config.js

Name		Name	Last commit message	Last commit date
.claude		.claude
src		src
.DS_Store		.DS_Store
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.js		tailwind.config.js
vite.config.js		vite.config.js

Name		Name	Last commit message	Last commit date
.claude		.claude
src		src
.DS_Store		.DS_Store
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.js		tailwind.config.js
vite.config.js		vite.config.js

Name		Name	Last commit message	Last commit date
.claude		.claude
src		src
.DS_Store		.DS_Store
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.js		tailwind.config.js
vite.config.js		vite.config.js

Latest commit

History

Folders and files

Repository files navigation

Infinite Research / 无限研究

Table of Contents / 目录

Features / 功能特性

Architecture / 架构

Component Overview / 组件概览

Message Processing Flow / 消息处理流程

Task Lifecycle / 任务生命周期

Prerequisites / 前置要求

Vane Installation / Vane 安装

Option A: Docker (Recommended / 推荐)

Option B: From Source / 从源码安装

Configure Ollama for Vane / 为 Vane 配置 Ollama

App Installation / 应用安装

Configuration / 配置

Environment Variables / 环境变量

Example .env / 示例

Usage / 使用方法

Tech Stack / 技术栈

License / 许可证

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Infinite Research / 无限研究

Table of Contents / 目录

Features / 功能特性

Architecture / 架构

Component Overview / 组件概览

Message Processing Flow / 消息处理流程

Task Lifecycle / 任务生命周期

Prerequisites / 前置要求

Vane Installation / Vane 安装

Option A: Docker (Recommended / 推荐)

Option B: From Source / 从源码安装

Configure Ollama for Vane / 为 Vane 配置 Ollama

App Installation / 应用安装

Configuration / 配置

Environment Variables / 环境变量

Example .env / 示例

Usage / 使用方法

Tech Stack / 技术栈

License / 许可证

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Infinite Research / 无限研究

Table of Contents / 目录

Features / 功能特性

Architecture / 架构

Component Overview / 组件概览

Message Processing Flow / 消息处理流程

Task Lifecycle / 任务生命周期

Prerequisites / 前置要求

Vane Installation / Vane 安装

Option A: Docker (Recommended / 推荐)

Option B: From Source / 从源码安装

Configure Ollama for Vane / 为 Vane 配置 Ollama

Example `.env` / 示例

Example `.env` / 示例

Example `.env` / 示例

Example `.env` / 示例