This repo implements a production-ready, scalable Retrieval Augmented Generation (RAG)-powered LLM-based Open Generative (or Extractive) context-aware Question-Answering (QA) App that:

• Takes as input a new query (or question);
• Implements vector similarity search within the embedding space by seeking relevant contexts corresponding to the incoming query in the vector database;
• Passes the relevant contexts as well as the input query to LLM;
• LLM then produces the answer to the input query while being aware of the relevant contexts related to the requested query.

This project also includes Fine-tuning a 20B parameters Large Language Model (LLM) in a multi-GPU cluster environment by leveraging the distributed training paradigm. Moreover, this repo develops scalable major ML workloads for contexts (load, embed, and index the contexts in the vector database) across multiple workers with different compute resources and serves the LLM App in a highly robust and scalable manner.

The below diagram shows the architectural design of this RAG-powered LLM App:

• Python
• Streamlit
• PEFT (for Parameter-Efficient Fine-Tuning)
• Accelerate
• Ray (for distributed LLM Fine-Tuning)
• Datasets
• Transformers
• PyTorch
• Numpy
• Scikit-Learn
• Deta (To access Deta Vector Database)
• LangChain
• FastAPI (To serve production-ready LLM App)

Squad dataset is used to fine-tune Eleuther AI's GPT-Neo 20B LLM model, which comprises Title, Question, Answer, and Context for each of the 98.2k dataset IDs.

• The Fine-Tuning process for GPT-Neo LLM model can be found in finetune.py file.
• The code to create RAG-powered LLM Agent for QA task can be seen in qa_agent.py file.
• To build the agent as production-ready API for QA task, it's worth delving deep into serve.py file.
• To seek prospects of using Streamlit to deploy the LLM app, head to streamlit.py file.
• All hyperparameters to control fine-tuning of the model are provided in the given config.py file.

To learn more about how to use this LLM RAG-powered QA App, consider watching the following video: