Socratic RAG Agent

A research topic exploration tool powered by Azure OpenAI and vector similarity search.

Table of Contents

1. Running Instructions
2. Technical Details
   • Topic Data Pipeline
   • Database & Embeddings
   • Topic Agent & Search
   • Chat Manager
   • User Interface
   • Cloud Infrastructure

1. Running Instructions

Prerequisites

• Python 3.8 or higher
• Azure CLI. Install from https://learn.microsoft.com/en-us/cli/azure/install-azure-cli
• Azure subscription with OpenAI access
• SSL Certificates for Azure PostgreSQL: Make sure you have installed the recommended SSL certificates to secure connections to our dedicated Azure PostgreSQL server. These certificates adhere to general industry standards and are not specific to this project. For detailed instructions, refer to the Azure PostgreSQL SSL/TLS Documentation.
• Azure Subscription Access: To utilize our Azure Subscription ID 0bdb7994-618e-43ab-9dc4-e5510263d104, you must obtain access through Professor Lew Lefton. Please request access by visiting the Cloudbank Billing Account Access portal or by contacting Professor Lew Lefton directly at lew.lefton@gatech.edu.

a. Azure Authentication

# Login to Azure - Be sure to use the account from cloudbank.org
az login

# Set the correct subscription
az account set --subscription 0bdb7994-618e-43ab-9dc4-e5510263d104

b. Create a Virtual Environment and Install Dependencies

python3 -m venv
source venv/bin/activate
pip install -r requirements.txt

c. Run the Application

streamlit run src/ui/streamlit_app.py

The application will be available at http://localhost:8501

2. Technical Details

a. Topic Data Pipeline

The system fetches research topics from the OpenAlex API using cursor-based pagination:

• Data validation and cleaning of topics and keywords
• Automatic keyword extraction and normalization
• JSON storage of raw data and keyword mappings

b. Database & Embeddings

PostgreSQL database with pgvector extension for similarity search:

• Topics table storing basic research topic information
• Keywords table with 768-dimensional SciBERT embeddings
• Many-to-many relationship table (topic_keywords)
• IVFFlat index for efficient vector similarity search
• Batch processing for embedding generation using SciBERT

c. Topic Agent & Search

Intelligent topic search system with:

• Multi-turn query understanding using conversation history
• Query rewriting for better context preservation
• Vector similarity search using pgvector extension
• Topic exclusion mechanism to avoid repetition
• Ranking based on keyword matches and similarity scores

d. Chat Manager

Manages the conversation flow with:

• Azure OpenAI usage for natural language processing
• Fallback mechanisms for API failures
• Conversation state handling

e. User Interface

Streamlit-based chat interface providing real-time interaction with the topic search system.

f. Cloud Infrastructure

All services are hosted in Azure under the OKN Project Cloudbank subscription:

• Resource Group: tikabox
• Services:
  • Azure PostgreSQL: Stores topics, keywords, and embeddings
  • Azure Key Vault: Manages service credentials and connection strings
  • Azure OpenAI Service: GPT-4 deployment for query processing and response generation
• Authentication: Managed through Azure DefaultAzureCredential