This project presents a multi-agent LLM-powered chatbot system capable of translating natural language queries into efficient SQL (and NoSQL) commands. Designed for non-technical users, it simplifies data access through an intuitive interface, powered by the latest advancements in language models and schema reasoning.

Built using LangChain, LangGraph, Streamlit, FastAPI, MongoDB, and Groq-hosted LLaMA 3.1.

• 🔍 Natural Language to Query: Converts user questions into executable MongoDB pipeline or SQL queries.
• 📊 Auto Visualization: Generates charts automatically for visual queries.
• 🧠 Schema-aware Prompting: Dynamically optimizes schema tokens to fit within LLM context.
• 🔁 Error-resilient Execution: Incorporates query validation, execution tracing, and iterative refinement.
• 💬 Interactive Chatbot: Built using Streamlit with real-time backend orchestration using LangGraph.

1. Clone the repository:

   git clone https://github.com/charangajjala/ai-project.git
   cd Query_GenAI

2. Install dependencies:

   pip install -r requirements.txt

3. Start the backend API server:

   uvicorn backend.main:app --reload

4. Start the Streamlit frontend:

   streamlit run frontend/app.py

• LangGraph Agents: Orchestrate logic for schema pruning, query generation, and error handling.
• LLM Model: LLaMA 3.1 (70B) via Groq API.
• Database: MongoDB Atlas — Sample Analytics Dataset.

The system is built using a modular multi-agent framework, where each component is designed to handle a specific function in the Text-to-SQL pipeline. The core workflow is orchestrated using LangGraph, enabling dynamic decision-making through routing logic.

• Extracts and prunes relevant schema details (tables, columns, data types) based on user queries.
• Reduces token count to stay within the LLM context window.
• Provides schema-aware few-shot prompts for accurate query generation.

• Converts natural language to SQL/MongoDB queries using LLaMA 3.1.
• Uses iterative prompting and few-shot examples.
• Validates and re-generates queries using error feedback if execution fails.

• Handles queries requesting plots or charts.
• Generates MongoDB queries and then Python code for plotting (e.g., bar, line charts).
• Returns visualizations rendered in real-time using Streamlit.

• Acts as the decision engine.
• Routes queries to either the query execution path or the visualization path.
• Falls back to error handling if inputs are incomplete or ambiguous.

1. Natural Language Input: User submits a query via chatbot UI.
2. Intent Detection: LangGraph router determines whether it's a data retrieval or visualization request.
3. Schema Optimization: Schema Agent fetches only necessary tables/columns.
4. Query Generation: LLM generates SQL/MongoDB queries with in-context examples.
5. Validation: Syntax and execution errors are caught, refined, and re-executed.
6. Output: Query results or plots are returned to the user.

This design emphasizes:

• 💡 Efficient token usage (schema pruning),
• 🔁 Iterative improvement (error-aware prompting),
• 📈 Visualization-first UX (interpretable results for non-technical users),
• 🧩 Scalability across SQL and NoSQL databases.

📊 Achieved: EM = 60%, EX = 80%, VES = 74%

Charan Gajjala Chenchu

Divija Kalluri

1. QueryGPT by Uber
2. Text-to-SQL with LLMs - arXiv 2023
3. LangGraph Documentation

This project is licensed under the MIT License.