ChatCapitalHumain

Overview

ChatCapitalHumain is an LLM-powered agent-based chatbot designed to analyze and provide insights on student survey data from Acadian schools in Atlantic Canada. The application connects to a PostgreSQL database containing questionnaire responses collected between 2004 and 2019 across seven schools in the region.

https://chatcapitalhumain.ca/ to interact with the system directly. The sidebar provides access to example questions and the database schema to help formulate effective queries.

Features

Database Schema

The application is built around a comprehensive database containing:

• Student responses to 52 questions across five categories (General Questions, Socio-Demographic Information, Post-Secondary Education, Job Market, Employment Search)
• Data from seven Acadian schools over a 15-year period (2004-2019)
• Aggregated responses by school, year, questionnaire, and gender

Dual Agent Architecture

The system implements two distinct agent architectures:

1. Single Agent (simple):

   • Direct database query processing with the LangChain ReAct pattern
   • Streamlined question-to-SQL flow for straightforward queries
   • Memory persistence across sessions via LangGraph's MemorySaver

2. Multi-Agent with Human-in-the-Loop :

   • LangGraph-powered swarm workflow with specialized agent nodes
   • Human feedback integration during the query formulation phase with Human-in-the-Loop
   • Transparent workflow visualization of agent handoffs
   • Structured outputs ensuring data quality and accuracy

Vector Search for Questions

• Pre-populated vector database containing example queries to help users understand what types of questions can be answered

Data Visualization

• Automated visualization of query results using Plotly in the Multi-Agent Architecture
• Interactive charts and graphs representing temporal trends and comparisons

Key Features

• Human-in-the-Loop: Expert validation of query interpretations before execution
• Login System: User authentication to save conversation history
• Schema Reference: Comprehensive database schema available in the sidebar
• Model Selection: Choice between different LLM models (Claude, GPT, Gemini)
• Memory Persistence: Conversations maintained across sessions
• Evaluation Framework: Benchmarking different models on question answering tasks

Technical Stack

• Frontend: Streamlit
• Database: PostgreSQL
• LLM Integration: LangChain, LangGraph
• Vector Search: FAISS
• Visualization: Plotly
• Authentication: Streamlit authentication for user management