奇美拉
具有RAG、网络搜索和持久内存的完全本地AI助手。
完全在您的硬件上运行。没有API密钥。没有云。没有遥测。
奇美拉是什么?
奇美拉连接 LM 工作室 通过 模型上下文协议。您当地的法学硕士获得了真正的能力:
知识与搜索
- 通过pgvector进行语义文档搜索
- 跨11个引擎的网络搜索(SearXNG)
- 通过Context7获取实时库文档
记忆与互动
- 具有语义回忆的会话记忆
- 持久知识图
- 文件系统访问(读/写/搜索)
文档处理
- 上传并索引PDF、DOCX、TXT、MD
- 使用TEI批量嵌入
- 余弦相似度排名
愿景与分析
- 屏幕截图与分析
- PDF文本提取
- 多模态视觉模型
> [!重要]
> 所有处理都在本地进行。文档嵌入在您的计算机上,存储在您的数据库中,永远不会离开您的网络。
建筑```
┌─────────────────────────────────┐
│ LM Studio (Host) │
│ Qwen 3.5 · localhost:1234 │
└───────────────┬─────────────────┘
│ MCP (stdio)
┌────────────────────────┼────────────────────────┐
│ │ │
┌────────┴────────┐ ┌──────────┴──────────┐ ┌───────┴───────┐
│ chimera-rag │ │ searxng · filesys │ │ memory · pdf │
│ (custom MCP) │ │ context7 · screenshot│ │ (stdlib MCP) │
└────────┬────────┘ └──────────┬──────────┘ └───────────────┘
│ │
▼ ▼
┌──────────────────────────────────────────────────────────────────┐
│ Docker Network │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ PostgreSQL │ │ TEI │ │ SearXNG │ │
│ │ + pgvector │ │ embeddings │ │ meta-search │ │
│ │ port 5432 │ │ port 8001 │ │ port 8888 │ │
│ └──────┬───────┘ └──────┬───────┘ └──────┬───────┘ │
│ └───────────────────┼───────────────────┘ │
│ ┌──────┴──────┐ │
│ │ RAG Server │ │
│ │ FastAPI │ │
│ │ port 8080 │ │
│ └─────────────┘ │
└──────────────────────────────────────────────────────────────────┘
back to top
## 入门指南
### 先决条件
|要求|最低要求|建议|
|-------------|---------|-------------|
| **图形处理器** |8GB VRAM | 16GB VRAM|
| **随机存取存储器** |16GB | 64GB|
| **软件** | Docker 桌面版, LM 工作室, Node.js 18+ | — |
### 第一步——克隆并启动Docker```bash
git clone https://github.com/figgiee/chimera.git
cd chimera/rag-setup
docker compose up -d
> [!提示]
> 首次启动需要大约3GB的Docker镜像,需要2-3分钟。后续启动是即时的。
等待健康检查:```bash
docker compose ps # all 4 services should show "healthy"
### 第二步——启动LM Studio
1. 打开LM Studio并加载模型
1. 在端口上启动服务器 `1234`
1. 复制 `mcp.json` 到LM Studio配置(请参阅 配置)
### 步骤3——将Docker连接到LM Studio
LM Studio绑定到 `localhost` 只有。Docker容器需要一个端口代理才能访问它:```powershell
# Run once in Admin PowerShell — persists across reboots
netsh interface portproxy add v4tov4 listenport=1234 listenaddress=0.0.0.0 connectport=1234 connectaddress=127.0.0.1
步骤4——验证```bash
curl http://localhost:8080/health
``````json
{
"status": "ok",
"services": {
"database": "ok",
"embeddings": "ok",
"search": "ok",
"llm_studio": "ok"
}
}
> \[!注意\]
> 如果 `llm_studio` 显示错误,请确保LM Studio正在运行并且端口代理已设置(步骤3)。
back to top
## MCP工具
奇美拉提供 **7台MCP服务器** 通过LM Studio的插件系统,您可以访问LLM工具。
### 奇美拉抹布-- *定制RAG管道*
|工具|说明|
|------|-------------|
| `search_documents` |具有可配置相似性阈值的跨索引文档的语义向量搜索|
| `upload_document` |通过自动分块和嵌入上传和索引文档(PDF、DOCX、TXT、MD)|
| `list_documents` |列出所有带有ID、文件名和内容预览的索引文档|
| `delete_document` |从知识库中删除文档及其所有嵌入|
| `store_conversation` |通过语义嵌入保存对话回合以用于长期记忆|
| `recall_conversation` |使用语义相似性搜索过去的对话|
| `rag_health` |检查所有RAG管道服务的健康状况|
### 外部MCP服务器
|服务器|源|功能|
|--------|--------|-------------|
| **搜索引擎** | mcp searxng |跨11个引擎(DuckDuckGo、谷歌、必应、维基百科、arXiv、GitHub、Brave等)的尊重隐私的元搜索|
| **文件系统** | @模型上下文协议/服务器文件系统 |跨配置的目录读取、写入、搜索和管理文件|
| **记忆** | @模型上下文协议/服务器内存 |跨会话实体和关系的持久知识图|
| **pdf阅读器** | mcp pdf阅读器 |高级PDF文本提取和结构化阅读|
| **截图** | mcp屏幕截图 |用于视觉模型分析的屏幕截图|
| **上下文7** | @upush/context7 mcp |按需查找最新的图书馆文档|
back to top
## Docker服务
Service · Image · Port · Description
chimera-rag-server · python:3.11-slim (custom) · 8080 · FastAPI orchestrator — handles document upload, search, conversation memory, and LM Studio integration
chimera-postgres · pgvector/pgvector:pg15 · 5432 · PostgreSQL with pgvector extension — stores 384-dimensional embeddings with cosine similarity search
chimera-tei · ghcr.io/huggingface/text-embeddings-inference · 8001 · HuggingFace Text Embeddings Inference — runs all-MiniLM-L6-v2 for fast vectorization
chimera-searxng · searxng/searxng:latest · 8888 · Privacy-respecting meta-search engine — aggregates 11 search backends with no API keys
back to top
## API 参考
GET /health — Service health check```bash
curl localhost:8080/health
``````json
{
"status": "ok",
"timestamp": "2026-03-11T07:45:09.469968",
"services": {
"database": "ok",
"embeddings": "ok",
"search": "ok",
"llm_studio": "ok"
}
}
POST /api/search — Semantic document search```bash
curl -X POST localhost:8080/api/search \
-H "Content-Type: application/json" \
-d '{"query": "machine learning", "limit": 5, "threshold": 0.3}'
|参数|类型|默认值|说明|
|-----------|------|---------|-------------|
| `query` |字符串| *必需的* |搜索查询文本|
| `limit` |int| `5` |返回的最大结果|
| `threshold` |浮子| `0.3` |最小余弦相似度(0-1)|
POST /api/documents/upload — Upload and index a document```bash
curl -X POST localhost:8080/api/documents/upload -F "file=@paper.pdf"
支持: .pdf, .docx, .txt, .md
文档会自动分块,通过TEI以32个为一批嵌入,并存储在pgvector中。
GET /api/documents — List indexed documents```bash
curl localhost:8080/api/documents
返回所有文档及其ID、文件名、上传时间戳和内容预览。
DELETE /api/documents/{id} — Delete a document```bash
curl -X DELETE localhost:8080/api/documents/{document_id}
从数据库中删除文档和所有关联的嵌入。
POST /api/conversations/store — Store a conversation turn```bash
curl -X POST localhost:8080/api/conversations/store \
-H "Content-Type: application/json" \
-d '{"conversation_id": "research-1", "role": "user", "content": "What is RAG?"}'
|参数|类型|说明|
|-----------|------|-------------|
| `conversation_id` |string |唯一对话标识符|
| `role` |字符串| `user` 或 `assistant` |
| `content` |string |要嵌入和存储的消息内容|
POST /api/conversations/recall — Recall past conversations```bash
curl -X POST localhost:8080/api/conversations/recall \
-H "Content-Type: application/json" \
-d '{"query": "RAG architecture", "limit": 5}'
| 参数 | 类型 | 默认值 | 说明 |
|---|---|---|---|
query |
字符串 | 必需的 | 语义搜索查询 |
conversation_id |
字符串 | null |
仅限于特定对话 |
limit |
int | 5 |
最大结果 |
back to top
配置
MCP服务器配置
将此复制到LM Studio MCP配置(~/.lmstudio/mcp.json),将路径替换为您的安装位置:
mcp.json```json
{
"mcpServers": {
"chimera-rag": {
"command": "node",
"args": ["/mcp-chimera-rag/index.js"],
"env": { "RAG_SERVER_URL": "http://localhost:8080" }
},
"searxng-search": {
"command": "node",
"args": ["/mcp-searxng/dist/index.js"],
"env": { "SEARXNG_URL": "http://localhost:8888" }
},
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "~/Desktop", "~/Documents", "~/Downloads"]
},
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"]
},
"pdf-reader": {
"command": "node",
"args": ["/mcp-pdf-reader/dist/index.js"]
},
"context7": {
"command": "npx",
"args": ["-y", "@upstash/context7-mcp@latest"]
},
"screenshot": {
"command": "node",
"args": ["/mcp-screenshot/dist/index.js"]
}
}
}
### 系统提示
Chimera的推荐系统提示在设计上是最小的——小型模型在简洁的说明下表现更好:
Optimized system prompt (~120 tokens)```
You are Chimera, a local AI assistant. Use your tools proactively. Be concise — lead with the answer.
Images, text, and code shared in chat are already in your context. Respond directly.
For dates, versions, current events, or anything time-sensitive: web_search first, then answer.
For questions that might relate to past sessions: recall_conversation to check, then answer.
After significant exchanges (research, analysis, problem-solving): store_conversation with a descriptive conversation_id.
If a tool fails, try an alternative approach. If a file path is given, read it immediately.
推荐模型
| 用例 | 型号 | VRAM | 速度 |
|---|---|---|---|
| 每日驱动程序 | Qwen 3.5-9B Q4_K_M | ~7 GB | ~93 tok/s |
| 繁重任务 | Qwen 3.5-35B-A3B Q4_K_M | ~14 GB | ~18 tok/s |
| 低VRAM | 3.5-9B Q3_K_M | ~5 GB | ~100 tok/s |
LM工作室设置(适用于35B型号)
| 设置 | 值 | 注释 |
|---|---|---|
| 上下文长度 | 16384 | 越高=VRAM越多,速度越慢 |
| GPU卸载 | 25 | 层卸载到GPU |
| CPU线程数 | 24 | 为操作系统留出空间 |
| 并发预测 | 1 | 避免上下文分割 |
back to top
项目结构```
chimera/
├── rag-setup/ # Docker infrastructure + RAG server
│ ├── docker-compose.yml # 4-service orchestration
│ ├── Dockerfile # RAG server image (Python 3.11)
│ ├── app.py # FastAPI application
│ ├── db.py # SQLAlchemy models + pgvector ops
│ ├── embeddings.py # TEI client (embed + batch)
│ ├── search.py # Document search + SearXNG client
│ ├── llm.py # LM Studio health + model detection
│ ├── init.sql # pgvector extension bootstrap
│ ├── requirements.txt # Python dependencies
│ ├── searxng-settings.yml # 11 search engine configs
│ ├── documents/ # Uploaded document storage
│ └── scripts/
│ ├── manage.sh # Service management CLI
│ └── ingest-documents.py # Batch document processor
│
├── mcp-chimera-rag/ # Custom MCP server
│ ├── index.js # 7 tool definitions + handlers
│ └── package.json
│
├── .env.example # Environment template
├── .gitignore
├── LICENSE
└── README.md
back to top
## 管理```bash
cd rag-setup
./scripts/manage.sh start # Start all services
./scripts/manage.sh stop # Stop all services
./scripts/manage.sh restart # Restart all services
./scripts/manage.sh rebuild # Rebuild RAG server after code changes
./scripts/manage.sh health # Check all service endpoints
./scripts/manage.sh logs # Tail all logs (or: logs rag-server)
./scripts/manage.sh db-backup # Backup PostgreSQL to ./backups/
./scripts/manage.sh reset # Nuclear reset — destroys all data
硬件要求
| 要求 | 最低要求 | 建议 |
|---|---|---|
| 随机存取存储器 | 16gb | 32Gb+ |
| GPU VRAM | 8 GB | 16 GB+ |
| 操作系统 | Windows 10、macOS 12或Linux | -- |
| 软件 | Docker桌面、LM Studio、Node.js 20+ | -- |
Chimera完全在本地运行-无需云,无需API密钥。性能取决于您在LM Studio中加载的模型。
back to top
建于 LM 工作室 · pg向量 · 瑟克斯NG · 拥抱脸TEI · 主控程序
MIT许可证
