A fully local AI assistant with Retrieval-Augmented Generation (RAG), document chat, and streaming responses powered by Ollama, ChromaDB, FastAPI, React, and Docker.
SecAI is a privacy-first local AI assistant that runs entirely on your own machine.
Developed as an internship project at Necurity Solutions, the goal of SecAI is to demonstrate how a production-style Retrieval-Augmented Generation (RAG) system can be built using open-source technologies while keeping all data and AI inference local.
No cloud APIs.
No subscription fees.
No external data sharing.
Users can upload their own documents, ask questions about them, and receive answers with source citations generated from the actual content of those documents.
- Local conversational AI using Mistral via Ollama
- Real-time streaming responses using Server-Sent Events (SSE)
- Persistent chat history
- Multiple chat threads
- Upload personal documents
- Semantic search using embeddings
- Context-aware responses
- Source citations included in answers
- Per-user document isolation
- DOCX
- XLSX
- TXT
- CSV
- JWT Authentication
- Password hashing
- Redis-based token revocation
- User-level document isolation
- No OpenAI API
- No cloud vector database
- No external AI services
- All data remains on your machine
Browser
│
▼
Nginx
│
├── React Frontend
│
└── FastAPI Backend
│
├── PostgreSQL
├── Redis
├── ChromaDB
└── Ollama (Host Machine)
- React
- Vite
- Tailwind CSS
- Axios
- FastAPI
- SQLAlchemy
- Alembic
- Pydantic
- Ollama
- Mistral
- LangChain
- ChromaDB
- Sentence Transformers
- all-MiniLM-L6-v2
- Docker
- Docker Compose
- Nginx
- PostgreSQL
- Redis
secai/
│
├── backend/
│ ├── app/
│ │ ├── core/
│ │ ├── models/
│ │ ├── schemas/
│ │ ├── services/
│ │ ├── routers/
│ │ └── utils/
│ │
│ ├── alembic/
│ ├── requirements.txt
│ └── Dockerfile
│
├── frontend/
│ ├── src/
│ ├── public/
│ ├── Dockerfile
│ └── package.json
│
├── nginx/
├── uploads/
├── docker-compose.yml
├── .env.example
└── README.md
Upload File
│
▼
Text Extraction
│
▼
Chunking
│
▼
Embedding Generation
│
▼
ChromaDB Storage
User Question
│
▼
Generate Query Embedding
│
▼
ChromaDB Similarity Search
│
▼
Retrieve Relevant Chunks
│
▼
Build Context
│
▼
Mistral via Ollama
│
▼
Stream Response + Citations
| File Type | Processing Library |
|---|---|
| pdfplumber | |
| DOCX | python-docx |
| XLSX | openpyxl |
| TXT | Native Reader |
| CSV | Python csv |
Install the following:
- Docker Desktop
- Git
- Ollama
curl -fsSL https://ollama.ai/install.sh | shDownload and install from:
Pull the default model:
ollama pull mistralVerify installation:
ollama run mistralgit clone https://github.com/YOUR_USERNAME/secai.git
cd secaiA complete .env.example file is included in the repository.
Copy it to .env and adjust values if required for your environment.
Build and start all services:
docker compose up --build| Service | Port |
|---|---|
| Frontend | 80 |
| FastAPI Backend | 8000 |
| PostgreSQL | 5432 |
| Redis | 6379 |
| ChromaDB | 8000 |
| Ollama | 11434 |
POST /api/auth/registerPOST /api/auth/loginPOST /api/auth/logoutGET /api/auth/mePOST /api/files/uploadGET /api/filesGET /api/files/{id}/statusDELETE /api/files/{id}POST /api/chat/threadsGET /api/chat/threadsDELETE /api/chat/threads/{id}GET /api/chat/threads/{id}/messagesGET /api/chat/threads/{id}/streamEach user receives an isolated ChromaDB collection:
user_{user_id}
Every document and chat operation is ownership-validated before access is granted.
This ensures users cannot access another user's documents, embeddings, or chat history.
| Metric | Target |
|---|---|
| First Token Latency | < 5 seconds |
| API Response Time | < 200 ms |
| 10 Page PDF Processing | < 30 seconds |
| Maximum Upload Size | 50 MB |
- Hybrid search combining vector and keyword retrieval
- Conversation memory
- OCR support for scanned PDFs
- Document summarization
- Knowledge graph integration
- Local model switching
- GPU acceleration options
- Advanced citation ranking
- Multi-document retrieval optimization
This project was developed as part of an internship at Necurity Solutions Network Security Private Ltd.
The project explores modern AI application architecture including local LLM deployment, Retrieval-Augmented Generation, vector databases, semantic search, document processing, and real-time streaming systems.
Subash Venkat