Context-Aware RAG
100% local, privacy-preserving RAG that maps territory before searching
Standard RAG systems have a massive blind spot: they search for terms the LLM doesn’t actually understand yet. Context-Aware RAG fixes this by mapping the territory before searching the database — all running 100% locally with zero API keys and zero data leaks.
Key Features:
- Local Tech Stack: Powered by Ollama (llama3.2) and ChromaDB — perfect for enterprise or highly sensitive personal documents
- Scanned PDF Mastery: Integrated pymupdf4llm with Tesseract OCR for image-only or scanned PDFs
- Smart Knowledge Compaction: Auto-deduplicates and compresses the knowledge.md index once it hits ~256K characters
- Domain Rules (Expert Mode): Extracts “rules of the game” from your docs — how you phrase queries, how you reason, how you want answers structured
- Sibling Context: Cross-Encoder reranker finds top 5 chunks, then grabs adjacent chunks for continuous context flow instead of scattered data
- Citations That Actually Work: Structured JSON answers with exact filename and quote — no hallucinations
- Bilingual OCR: Optimized for English + Hindi, especially useful for Indian regulatory documents
- Streamlit UI: Clean built-in frontend for document upload, vector chunk inspection, and chat