03-rag-systems

Module 03 · RAG Systems

Build Retrieval-Augmented Generation pipelines that ground AI responses in your enterprise documents — eliminating hallucinations and keeping answers current without retraining.

Files

File	Description
full_rag_pipeline.py	End-to-end pipeline: ingest → embed → retrieve → generate

Architecture

Documents → Chunker → Embeddings (HuggingFace) → ChromaDB
                                                        ↓
User Query → Embeddings → Retrieval → Context Builder → LLM (Grok/OpenRouter) → Answer

Quick Start

# Run the full demo (creates local chroma_db/)
python 03-rag-systems/full_rag_pipeline.py

Key Parameters (via .env)

Variable	Default	Description
EMBED_MODEL	all-MiniLM-L6-v2	Embedding model
CHROMA_PERSIST_DIR	./chroma_db	Vector store location
RAG_CHUNK_SIZE	500	Words per chunk
RAG_CHUNK_OVERLAP	50	Overlap between chunks
RAG_TOP_K	4	Chunks retrieved per query
OPENROUTER_RAG_MODEL	mistralai/mistral-small	LLM for generation

Scaling to Production

• Swap ChromaDB → Pinecone / pgvector for multi-tenant scale
• Add reranking with cross-encoder/ms-marco-MiniLM-L-6-v2
• Implement hybrid search (dense + BM25) for better recall
• Add metadata filters to scope retrieval by department/date