Retrieval-Augmented Generative AI with Fine-Tuned Mistral-7B Model and llama

https://colab.research.google.com/drive/1FgUgJKY7HVFF8_YJumjwnsEiYdTwLGQM?usp=sharing

Overview

This project combines Retrieval-Augmented Generation (RAG) with a fine-tuned generative AI model to create a personalized conversational assistant. It dynamically retrieves context from uploaded documents and generates accurate, context-aware responses. The assistant, named KyeGPT, serves as a virtual representation of me, answering questions about my professional background, skills, and experiences.

Key Features

1. Semantic Search for Contextual Retrieval:

   • Utilizes a dense vector embedding model (BAAI/bge-small-en-v1.5) for semantic search, ensuring contextually relevant documents are retrieved based on the meaning of queries rather than keyword matching.

2. Fine-Tuned Language Model:

   • Fine-tuned Mistral-7B-Instruct-v0.2-GPTQ model using PEFT (Parameter-Efficient Fine-Tuning) to generate responses tailored to my resume, cover letter, and project experiences.

3. Dynamic Prompt Engineering:

   • Integrates retrieved content into prompts, ensuring the generated responses are enriched with accurate and contextually relevant information.

4. Deployment-Ready Packaging:

   • Saves the fine-tuned model and tokenizer for deployment, enabling the conversational assistant to be hosted on platforms such as AWS EC2.

   
   
   
   

How It Works

1. Document Processing:

   • Upload documents (e.g., resumes, cover letters) to be indexed.

   • Chunk the documents into smaller sections for efficient retrieval.

2. Retrieval-Augmented Generation:

   • Retrieve the top 3 most relevant document chunks based on a similarity threshold.

   • Pass the retrieved content to the fine-tuned model as additional context for query-based text generation.

3. Response Generation:

   • Generates accurate and concise answers to queries using the integrated retrieval and generation pipeline.

Technical Stack

• Language Model: Mistral-7B-Instruct-v0.2-GPTQ

• Embedding Model: BAAI/bge-small-en-v1.5

• Libraries:

  • LlamaIndex: For document indexing and retrieval.

  • Transformers: For model inference and tokenization.

  • PEFT: For fine-tuning the language model efficiently.

• Tools and Frameworks:

  • Python, Flask (for API), and AWS (for deployment).

Innovations

1. Seamless Integration of RAG with Generative AI:

   • Combines the strengths of retrieval systems with generative models to enhance response accuracy and contextual relevance.

2. Parameter-Efficient Fine-Tuning:

   • Optimized training using PEFT to ensure computational efficiency while achieving high-quality output.

Sample Query and Response

• Query: "Why should I hire you?"

• Generated Response: "As a highly motivated individual with expertise in machine learning, data analytics, and AI, I bring a unique combination of technical skills, problem-solving ability, and teamwork. My journey from being a disciplined baseball player to excelling in data science has equipped me with resilience and adaptability to tackle complex challenges effectively."

Future Improvements

• Advanced Retrieval Techniques:

  • Explore graph-based approaches or hybrid retrieval models for better contextual understanding.

• User Interaction:

  • Develop a web-based GUI for easier interaction with the assistant.

• Scalability:

  • Deploy the system using containerization (e.g., Docker) for efficient scaling across multiple environments.