RAG: Retrieval-Augmented Generation System

This project implements a complete RAG (Retrieval-Augmented Generation) pipeline designed to "chat" with audio content. It processes audio data, transcribes it using AI, and creates a searchable vector database to answer user queries with high precision.

---

Project Milestones

Based on the development workflow:

• Audio Transcription: Using OpenAI Whisper to translate and transcribe .mp3 files.
• Data Chunking: Breaking down long video/audio files into smaller segments with preserved metadata.
• Vector Embeddings: Generating 1024-dimensional embeddings using the BGE-M3 model via Ollama.
• Database Management: Using Pandas for bulk loading and Joblib for efficient dataframe serialization.
• Semantic Search: Implementing Cosine Similarity to retrieve the most relevant video chunks for any user question.
• LLM Integration: Feeding retrieved context into a Large Language Model to generate human-like responses.

---

🛠️ Tech Stack

• LLM/Embeddings: Ollama (BGE-M3 & Llama 3)
• Transcription: OpenAI Whisper
• Data Handling: Pandas, NumPy, Joblib
• Vector Math: Scikit-Learn (Cosine Similarity)
• Environment: Anaconda Python 3.x

---

Getting Started

1. Prerequisites

Ensure you have Ollama installed and the embedding model pulled: ollama pull bge-m3

2. Project Pipeline

• Transcription: Convert video/audio to text chunks.
• Embedding: Run the script to generate vectors for each chunk.
• Saving: The dataframe is saved as embeddings.joblib for persistent storage.
• Querying: Run process_incoming.py to ask questions.

3. System Architecture

• User Query: The user asks a question via the terminal.
• Embedding: Run the script to generate vectors for each chunk.
• Retrieval: The system compares the query vector against the stored database (5,916 vectors).
• Augmentation: Top matching chunks are pulled and formatted into a prompt.
• Generation: The LLM reads the chunks and provides a factual answer based on the video content.