PolicyBrain is a production-ready AI system that answers complex policy and compliance questions using Retrieval-Augmented Generation (RAG), confidence scoring, and multi-step reasoning.
It is designed for domains where accuracy, traceability, and confidence estimation matter — such as healthcare, cloud compliance, and regulatory policy.
- Advanced RAG Pipeline
- Hybrid retrieval (vector + keyword)
- Multi-hop query reasoning
- Context-aware answer generation
- Confidence Scoring
- Explicit confidence estimation for each answer
- Low-confidence handling for vague or underspecified questions
- Policy-Safe Outputs
- Grounded responses strictly based on retrieved policy text
- Reduced hallucination risk
- Production-Ready API
- FastAPI backend
- Typed request/response schemas
- Automated tests
- Performance Optimized
- Cached embeddings
- Persistent FAISS vector index
User Query ↓ Query Rewriting ↓ Hybrid Retrieval (Vector Search + Keyword Search) ↓ Multi-Hop Reasoning ↓ Answer Generation ↓ Confidence Scoring ↓ Final Response (Answer + Confidence + Citations)
policybrain/ ├── api.py # FastAPI entry point ├── ingestion/ # Policy chunking & indexing ├── retrieval/ # Vector, keyword & hybrid search ├── reasoning/ # Answer + confidence generation ├── tests/ # API & regression tests ├── data/ # Policy documents ├── requirements.txt └── start.sh # Production startup script
- Pytest-based API tests
- Covers:
- Normal queries
- Vague / ambiguous queries
- Low-confidence responses
- Regression behavior
Run tests:
pytest
### ⚙️ Running Locally
pip install -r requirements.txt
uvicorn api:app --reload
## 🎯 Use Cases
Healthcare policy interpretation
Cloud compliance (HIPAA, HHS)
Internal policy Q&A
Regulatory decision support