Overview

This project implements a dual-function Retrieval-Augmented Generation (RAG) assistant that bridges the gap between complex theoretical research and practical implementation. It not only helps researchers and developers understand and implement ML/AI research papers with efficient CUDA optimization, but also breaks down complex mathematical topics into digestible, visually engaging animated explanations.

By combining paper comprehension, advanced mathematical summarization, and Manim-based animation with CUDA documentation retrieval, the system provides a comprehensive tool for both academic exploration and practical GPU programming.

Features

Scientific Paper & Math Processing:

Processes academic papers or math-heavy texts (PDF or plain text) to extract and preserve technical details and mathematical notation.
Uses advanced text chunking and embedding techniques to maintain context and isolate key concepts.
Generates concise summaries that break down complex math topics into step-by-step instructions suitable for visualization.
Manim-Based Animation Generation:

Automatically converts extracted mathematical concepts into detailed Manim animations.
Produces scene-by-scene breakdowns—for example, plotting functions, highlighting transformations, and visualizing derivations—to bring complex concepts to life in a style reminiscent of 3Blue1Brown.
Allows interactive adjustments of visualization parameters (such as domain, range, or color highlights).
CUDA Implementation Guidance:

Searches through CUDA documentation and NVIDIA best practices to provide targeted strategies for implementing and optimizing ML/AI algorithms on GPUs.
Utilizes vector search with FAISS to retrieve the most relevant documentation and presents contextually accurate solutions using LLM-generated responses.
LLM Integration:

Leverages Hugging Face's FLAN-T5 model for generating both context-aware paper summaries and detailed Manim scene instructions.
Ensures that responses are both theoretically sound and practically applicable for GPU programming and mathematical visualization.
Project Structure

graphql
Copy
Copyproject/
├── notebooks/
│ ├── 01_data_preparation.ipynb # PDF extraction, CUDA & math document processing
│ ├── 02_embeddings.ipynb # Generating vector embeddings for papers and CUDA docs
│ ├── 03_model_testing.ipynb # Testing response generation and animation pipelines
│ └── 04_quality_improvement.ipynb # Optimizing output quality for responses and animations
├── app.py # Streamlit web application integrating both functionalities
├── data/ # Datasets (e.g., paper PDFs, math text, annotated Manim prompts)
├── src/
│ ├── doc_processing/ # Code for processing papers and math content
│ ├── visualization/ # Manim pipeline and math visualization tools
│ ├── embeddings/ # Embedding generation and FAISS vector search
│ ├── models/ # LLM and response generation models
│ └── utils/ # Utility functions and helpers
├── requirements.txt # Project dependencies
└── Dockerfile # Container configuration for deployment