Agentic Physics Question Generator: An Agent-Driven RAG-Based Automatic Physics Problem Generation System

This project proposes an agent-driven Retrieval-Augmented Generation (Agentic RAG) technology, combined with the Qdrant vector database and reinforcement learning, to address the problems of lack of diversity in traditional physics problem generation methods and frequent errors in Large Language Models (LLMs). It achieves automatic generation of logically consistent and novel physics application problems, providing an innovative solution for the AI education field.

In STEM education, high-quality physics problem generation needs to balance concept coverage, logical consistency, and contextual value. However, traditional template-based methods lack diversity, and Large Language Model (LLM)-generated problems often have conceptual errors or inconsistent numerical values, making it difficult to meet educational needs.

The project innovatively integrates agent autonomous decision-making with RAG's fact-grounding capabilities through an iteratively optimized retrieve-evaluate-adjust loop: the agent constructs optimal vector queries, evaluates the relevance of Qdrant retrieval results, adjusts strategies, then integrates context to guide LLM in generating problems, improving the accuracy of knowledge support.

A reinforcement learning mechanism is introduced, using a Q-function to evaluate problem quality. Reward signals are obtained based on three dimensions: solvability (consistent parameters, clear problem-solving paths), novelty (innovation in context/parameters/thinking), and difficulty balance. Users can configure weight ratios to adapt to different needs (e.g., basic training or competition preparation).

The project code is modular, including modules for datasets, LLM core logic, novelty scoring, etc. Users start the system via main.py and configure settings in config.yaml: topics (add physics topics and formulas/contexts/difficulty levels), generation control (number of problems, dataset persistence), model selection (OpenAI models and API keys), and reinforcement learning parameters (Monte Carlo updates, learning rate).

The project can be applied to personalized practice (teaching according to individual needs), question bank expansion (rapid and high-quality scaling), adaptive testing (difficulty adaptation), and teaching research (analyzing problem distribution and student misconceptions), empowering the intelligence of AI education.

The project demonstrates that the Agentic RAG model can be extended to complex knowledge-intensive tasks. Multi-objective optimization provides tools for quality control, and structured domain knowledge is worth learning from. In the future, capabilities such as multi-modal generation, reverse optimization via answer analysis, and cross-disciplinary problem generation can be expanded.