A privacy-first, state-machine driven Retrieval-Augmented Generation (RAG) ecosystem for high-accuracy document intelligence.
Veritas AI is a production-oriented Agentic RAG platform engineered for industries where accuracy and privacy are non-negotiable (e.g., Medical, Fintech, Legal). Unlike traditional "Linear RAG" pipelines that often suffer from context loss and hallucinations, Veritas utilizes a Graph-based State Machine to treat retrieval as an iterative reasoning process. By combining local embedding generation with high-speed inference, the platform ensures that sensitive data remains traceable, grounded, and secure.
The core of Veritas is a deterministic graph that manages the flow of information:
- Intelligent Router: Analyzes user intent to decide between direct synthesis or multi-step document retrieval based on state context.
- Evidence Grounding: A "Generator" node programmatically restricted to only use retrieved context, effectively eliminating off-topic hallucinations.
- State Persistence: Utilizes a structured state object to maintain context throughout complex, multi-turn reasoning loops.
- Local Embedding Ingestion: Veritas uses
SentenceTransformers(all-MiniLM-L6-v2) to generate embeddings on-premise. Sensitive document content is never transmitted to third-party embedding providers. - Stateless Inference: Leveraging Groq Cloud (Llama 3.3 70B) via optimized API layers to deliver sub-second inference speeds while maintaining strict data handling boundaries.
- Vector Search: Semantic search is powered by a containerized Qdrant instance, utilizing optimized HNSW indexing for high-speed similarity lookups across large-scale document corpuses.
| Layer | Technology | Key Usage |
|---|---|---|
| Backend API | FastAPI (Python 3.10) | High-performance asynchronous API gateway |
| AI Orchestration | LangGraph / LangChain | State-machine based agent logic and flow control |
| Vector Engine | Qdrant (Docker) | Low-latency semantic search and metadata filtering |
| Embeddings | SentenceTransformers | Local-first, privacy-preserving vectorization |
| LLM Inference | Groq (Llama 3.3 70B) | Ultra-fast token generation and reasoning |
| Frontend UI | Next.js 14 / Tailwind | Responsive, TypeScript-based dashboard |
| DevOps | Docker / Docker Compose | Reproducible, containerized deployment ecosystem |
- Automated Document Ingestion: PDF processing pipeline including semantic chunking, embedding generation, and vector storage
- Traceable Reasoning: Logs full reasoning path (Router → Retriever → Generator) for auditability
- Interactive Dashboards: Next.js interface for document management and chat interaction
- Developer-First API: OpenAPI (Swagger/ReDoc) documented endpoints
- Docker & Docker Compose
- Python 3.10+
- Node.js 18+
- Groq API Key
Launch the Qdrant vector database:
docker-compose up -dcd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
# Create .env file with:
# QDRANT_HOST=localhost
# QDRANT_PORT=6333
# GROQ_API_KEY=your_api_key
uvicorn app.main:app --reloadcd frontend
npm install
npm run dev- Multi-Vector Retrieval (Parent-document retrieval)
- Feedback Loop for response evaluation
- Self-correction reviewer node for validation
Negil Babu
Software Engineer | MSc Artificial Intelligence
Berlin, Germany
LinkedIn | GitHub | Portfolio
This project is licensed under the MIT License.