Math Tutor Agent: Development Summary

📋 Chain of thought | Fun facts:

• Leaked API key doesn't work anymore, generate yours on developer.wolframalpha.com

• I created an Agent with 2 tools: tool 1 - checks if a query is about math, and the 2nd is just a whole logic of math tutoring.

• Streamlit frontend, just open an .ipynb file and use it straight in browser. LINK to the video example!

• I use Gemma 2 9B instruct because it can generate Ukrainian quite well.

• The ChromaDB dataset doesn't even exist, but if you asked the agent the same question - it'll create a brand new database and give you back a previously saved answer. (for generation of more robust answers we can use a bigger model, but even the current one is great at math questions (though it can't solve image input))

  

• My goal was to use the small 3B Llama 3.1 for agent logic and use Gemma 2 9B for explanation (now it's Gemma 2 9B instruct for both)

• For equations, I use the Wolfram Simple API.

• I was planning to add a button to tasks generation on the sidebar (but didn't make it yet)

• I was planning to generate ZNO questions for ChromaDB + Image library (but didn't make it). The problem is with understanding images by LLMs and image generation for the answer.

• Validation of dataset isn't working without a dataset, because it's not a RAG but an Agent without a default dataset.

• I've planned to save questions, answers with explanation[s] and with pictures, and I made it.

• I use Ollama instead of vLLM because I'm working on a laptop with 4GB GPU AND I want to use as models as small as possible to understand how to use them properly.

📋 Overview

The project represents an intelligent system for solving mathematical problems using a combination of modern technologies and approaches.

✅ Implemented Components

1. Core Infrastructure

• Wolfram Alpha integration for precise calculations
• Result caching system using ChromaDB
• Bilingual support (Ukrainian/English)
• Web interface based on Streamlit

2. Problem Processing

• Automatic mathematical problem type detection
• Step-by-step explanation generation
• Result visualization through graphs and diagrams
• Solution storage and reuse

3. System Reliability

• User input validation
• Error handling at all levels
• Fallback mechanism for local calculations
• API key and connection verification

🔄 Technology Stack

• Core Components:

  • Wolfram Alpha API
  • ChromaDB
  • Streamlit
  • Ollama
  • Google Translator

• Supporting Tools:

  • PIL for image processing
  • JSON for data structuring
  • Requests for API calls

📈 Potential Improvements

1. Functionality Extension

• Implementation of task generator
• Addition of evaluation system
• Development of multi-agent architecture

2. Quality Improvements

• Knowledge base expansion
• Enhanced explanations
• Translation optimization

3. Technical Enhancements

• Caching optimization
• Error handling improvements
• Expanded visualization format support

📊 Conclusions

The project demonstrates good basic functionality for mathematical problem solving but requires further development in task generation and quality assessment. The core system components work reliably and efficiently, providing a solid foundation for future development.

Key Achievements:

1. Created a working prototype with basic functionality
2. Implemented a reliable mathematical problem processing system
3. Provided bilingual support and user-friendly interface

Development Directions:

1. Implementation of task generation
2. Development of evaluation system
3. Realization of multi-agent architecture

💡 Technical Implementation Details

RAG System

• Uses ChromaDB for efficient storage and retrieval

• Implements similarity search for cached solutions

  

• Maintains metadata for better search results

Translation Pipeline

• Bidirectional translation between Ukrainian and English
• Handles mathematical terminology accurately
• Preserves formatting during translation

Error Handling

• Multiple fallback mechanisms
• Graceful degradation when services are unavailable
• Comprehensive error messages in Ukrainian

Visualization

• Dynamic image generation through Wolfram Alpha
• Local image storage and management
• Responsive display in web interface

🔍 Code Quality

Strengths

• Well-structured modular design
• Comprehensive error handling
• Clear separation of concerns
• Type hints and documentation

Areas for Improvement

• Need for more unit tests
• Better configuration management
• Code optimization opportunities
• Enhanced logging system

📚 Documentation Status

Complete

• Basic setup instructions
• API documentation
• Configuration guide