LLM-RAG-Gemma_7b_it

Introduction

The repo is a demonstration of implementing RAG paradigm using an open-source LLM (Gemma-7B-Instruct). Gemma is Google's open source lightweight LLM model. The work also explores different approaches and techniques (advanced vs naive) to create an end to end RAG pipeline.

Components

• LLM - Gemma-7b-it
• Embedding Model - BAAI/bge-large-en-v1.5
• Vector DB - Pinecone
• Orchestrator - LangChain
• Re-ranker - Cohere
• Models Source - HuggingFace
• Run-time environment - Google Colab T4 GPU

Steps

• Load Documents
• Chunk text
• Create Embeddings using model - BAAI/bge-large-en-v1.5
• Index and load to Vector Store - Pinecone
• Load Quantized LLM model - Gemma-7b-it
• Build Question Answering Chain using LangChain
• RAG w/ Base Retriever - Similarity search with pre-defined Threshold and Top-k
• RAG w/ pre-retrieval (metadata filtering) and Re-ranker - Cohere