yanacuti1121 Production RAG system architecture từ "The Mother of AI Project" course.
Core principle
Build RAG the professional way: solid keyword search foundations first, then enhance with vectors — not AI-first approaches that skip search fundamentals.
Stack
FastAPI — API layer
PostgreSQL — metadata store
OpenSearch — BM25 keyword + vector hybrid search
Airflow — data pipeline orchestration
Docker Compose — full local stack
Architecture phases
- Infrastructure — Docker, FastAPI, PostgreSQL, OpenSearch, Airflow
- Data pipeline — automated fetch + parse (arXiv papers or any domain)
- BM25 keyword search — production search with filtering + relevance scoring
- Vector search — semantic embeddings layer on top of keyword
- Hybrid retrieval — combine BM25 + vector for best results
- Agentic RAG — agent loop with tool use on top of retrieval
Key patterns
# Hybrid search: keyword + semantic
results = hybrid_search(
query=user_query,
bm25_weight=0.4,
vector_weight=0.6,
top_k=10
)