CUDA Documentation QA System

This project implements a Question Answering (QA) system for CUDA documentation. It crawls the NVIDIA CUDA documentation, processes the data, stores it in a vector database, and uses advanced retrieval techniques to answer user queries.

Features

• Web crawling of NVIDIA CUDA documentation
• Advanced data chunking based on semantic similarity
• Vector embedding creation and storage in Milvus database
• Query expansion for improved retrieval
• Hybrid retrieval combining BM25 and BERT-based methods
• Question answering using a Language Model

Setup Instructions

Prerequisites

• Python 3.7+
• pip (Python package installer)

Installation

1. Clone the repository:
2. Create a virtual environment (optional but recommended):
3. Install the required dependencies:

Dependencies

The main dependencies for this project are:

• scrapy: For web crawling
• sentence-transformers: For text embeddings
• nltk: For natural language processing tasks
• rank_bm25: For BM25 retrieval
• torch and transformers: For working with transformer models
• streamlit: For creating web applications
• selenium and webdriver_manager: For web scraping
• pymilvus: For interacting with the Milvus vector database

For a complete list of dependencies, refer to the requirements.txt file.

Running the System

1. Ensure that you have a Milvus server running. Refer to the Milvus documentation for installation and setup instructions.

2. Run the main script: 3. The system will start by crawling the CUDA documentation, processing the data, and storing it in the Milvus database. This initial setup may take some time.

3. Once the setup is complete, you can start asking questions about CUDA. The system will provide answers based on the retrieved information.

4. To exit the system, type 'quit' when prompted for a question.

Project Structure

• main.py: The main script that orchestrates the entire process.
• crawler/web_crawler.py: Contains the web crawling logic.
• data_processing/chunking.py: Implements advanced data chunking techniques.
• data_processing/embedding.py: Handles the creation of vector embeddings.
• vector_db/milvus_db.py: Manages interactions with the Milvus database.
• retrieval/query_expansion.py: Implements query expansion techniques.
• retrieval/hybrid_retrieval.py: Contains the hybrid retrieval logic.
• qa/llm_qa.py: Manages the question answering process using a language model.

Customization

• You can adjust the embedding model by modifying the SentenceTransformer model in main.py.
• The depth of web crawling can be adjusted in the crawl_data function (currently set to 5 levels).
• The number of retrieved chunks for answering can be modified by changing the top_k parameter in the retrieve method call.

Troubleshooting

If you encounter any issues:

• Ensure all dependencies are correctly installed.
• Check that the Milvus server is running and accessible.
• Verify that you have a stable internet connection for web crawling and model downloads.

For any persistent problems, please open an issue in the GitHub repository.