Refining-a-LLM-using-RAG

The goal of this project is to create a system for refining Large Language Models (LLMs) using Retrieval-Augmented Generation (RAG). It implements a design assistant that leverages GPT-4 and ElasticSearch to provide contextually relevant design advice.

Prerequisites

• OpenAI API key
• ElasticSearch instance
• Kibana
• Python 3.8+

Installation

1. Clone the repository:

git clone https://github.com/MaxRondelli/Refining-a-LLM-using-RAG.git
cd Refining-a-LLM-using-RAG

2. Install required dependencies:

pip install -r requirements.txt

3. Create a .env file in the root directory with your credentials:

OPENAI_API_KEY=your_openai_api_key
ELASTIC_HOST=your_elasticsearch_host
CA_CERTS_PATH=ca_certs_path
ELASTIC_USERNAME=your_username
ELASTIC_PWD=your_password

Project's Structure

• webapp.py: Streamlit-based user interface and chat logic
• main.py: Core functionality for document retrieval and prompt generation
• indexer.py: Document processing and embedding generation
• elastic.py: ElasticSearch database configuration and operations

Project's Pipeline

Document Indexing

The system processes PDF documents through the following steps:

• Splits documents into manageable chunks.
• Generates embeddings for each chunk using OpenAI's embedding model.
• Stores the embeddings and text in ElasticSearch.

Query Processing

When a user submits a query:

• The system generates an embedding for the query.
• Searches for the 3 most relevant document chunks using k-NN search.
• Creates a prompt template combining the query and relevant context.

Response Generation

The system:

• Maintains conversation history for context.
• Uses GPT-4 to generate responses based on retrieved documents.
• Presents responses through a user-friendly interface.

Usage

1. Create the vector database running the

python3 elastic.py

Be sure the docker with elastic and kibare are on.

2. Index the new documents with

python3 indexer.py

3. Start the web application with

streamlit run webapp.py

Contributing

The project has been developed between me and Alessandro Borrelli.

Contributions are welcome! Feel free to submit PRs.