Anatomy of a Full RAG Application: Every Concept, One Self-Hosted Stack

Short summary

A detailed walkthrough of myRAG, a fully self-hosted RAG stack using FastAPI, React, Qdrant, PostgreSQL, and Neo4j. The pipeline covers document conversion via Docling, recursive chunking, dual indexing (dense embeddings + BM25 sparse vectors), RRF fusion, cross-encoder reranking, LLM triple extraction into a knowledge graph, and token-budget-aware context assembly. Includes real code snippets for each stage.

• •Self-hosted RAG with six Docker containers: app, frontend, Docling, Qdrant, Postgres, Neo4j
• •Hybrid search combines dense embeddings and BM25 sparse vectors with RRF fusion
• •Knowledge graph triples from LLM extraction handle relational questions that pure vector RAG misses

Generated with AI, which can make mistakes.