AI Courtroom: A Legal Simulation Education Platform Based on RAG and Multi-Provider LLMs

AI Courtroom is a full-stack legal simulation education platform integrating Next.js 15, FastAPI, LangChain, and MongoDB. It uses Retrieval-Augmented Generation (RAG) technology to build standardized case contexts, supports AI-driven case generation, witness interaction, courtroom debate, and intelligent adjudication. It aims to address the problem of limited and costly practical opportunities in legal education, providing an immersive simulation environment for law learners.

Legal education has long faced core challenges: students need plenty of practical opportunities to develop courtroom debate, evidence analysis, and legal reasoning skills, but real courtroom internship opportunities are limited and costly. The AI Courtroom project attempts to build a virtual courtroom simulation platform using generative AI technology, allowing learners to practice repeatedly in a safe and controlled environment. Additionally, it uses RAG technology to avoid large model hallucination issues and ensure content complies with legal norms.

Core functional modules include:

1. AI-driven case generation: Automatically generates complete case scenarios (factual background, dispute focus, applicable legal provisions) based on user-specified legal fields and dispute types, and assigns jurisdiction;
2. Witness interrogation system: AI-generated witness roles maintain testimony consistency, supports interactive interrogation, and records the process for review;
3. RAG-enhanced debate environment: Uses Sentence Transformers to generate vector embeddings, retrieves relevant precedents and legal provisions as context, ensuring counterarguments have legal basis;
4. Intelligent adjudication generation: Simulates judges to generate rulings and case strength analysis based on arguments and evidence quality.

The technical architecture adopts a front-end and back-end separation design:

• Frontend: Next.js15 (React Server Components for performance improvement), Tailwind CSS4, Framer Motion, Capacitor (native app packaging), Serwist (PWA offline support);
• Backend: FastAPI (asynchronous web framework), Beanie ODM (MongoDB operation), LangChain (RAG orchestration and multi-provider failover, supporting Groq and OpenRouter);
• Data layer: MongoDB stores user information, case data, etc., integrates Country State City API and automatically refreshes cache;
• Authentication and security: JWT Token authentication (email/password + Google OAuth), request ID tracking and structured logging.

RAG system implementation details:

1. Document splitting and embedding: Case documents are split into text chunks, and the all-MiniLM-L6-v2 model is used to generate 384-dimensional vector embeddings (lightweight, suitable for CPU inference);
2. Vector storage and retrieval: Embedded vectors are stored in MongoDB, similarity retrieval is performed using vector search, and relevant fragments are injected into LLM prompts as context;
3. Multi-provider failover: Automatically switches to OpenRouter when the Groq API is unavailable to ensure service continuity.

Educational value and application scenarios:

• Case law learning: Practice identifying legal issues, retrieving precedents, and building argument chains; instant feedback helps with iteration;
• Oral debate training: Simulates dynamic interactions in real court trials, cultivates on-the-spot response and logical expression skills;
• Cross-jurisdiction learning: Supports simulation of high courts in different countries/regions to understand different legal systems;
• Collaborative learning: Multiple students play different roles (plaintiff, defendant, witness) for adversarial simulation.

Deployment and operation considerations:

• Provides Docker Compose configuration to simplify development and deployment;
• For production environments, it is recommended to host the frontend on Vercel and the backend on Render (in line with Jamstack best practices);
• Private deployment: The MIT license allows modification and secondary development; multiple API keys (Groq, OpenRouter, etc.) need to be configured, and a secure key management mechanism must be established.

Limitations and improvement directions:

• Current limitations: Although AI-generated content is enhanced by RAG, it may still be inaccurate and cannot replace professional legal advice; it relies on external LLM APIs, and costs increase with user scale;
• Improvement directions: Support local deployment of open-source legal large models to reduce API costs; add more case libraries for different jurisdictions; develop course management functions for teachers; introduce multimodal support for evidence image and video analysis.