MiroFish is a cutting-edge swarm intelligence engine designed to predict a wide range of outcomes by simulating complex systems. Its core function is to create high-fidelity digital worlds populated by numerous intelligent agents that interact and evolve, mirroring real-world scenarios. The project's primary goal is to provide a platform for forecasting, experimentation, and understanding the potential futures of various systems, from policy decisions to fictional narratives.

At its heart, MiroFish utilizes multi-agent technology. It takes "seed information" from the real world, such as news articles, policy drafts, or financial data, and uses this to construct a parallel digital environment. Within this environment, thousands of independent agents, each with their own personalities, memories, and behavioral logic, interact freely. Users can then manipulate variables within this simulated world, allowing them to observe and analyze the potential consequences of different actions. The engine aims to provide a "God's-eye view" of these simulations, enabling users to predict future trends and make informed decisions.

The project's main features revolve around its ability to ingest data, build simulations, and generate insightful reports. The workflow begins with "graph construction," which involves extracting information from real-world sources, injecting individual and collective memories into the agents, and building a knowledge graph using GraphRAG (Graph-based Retrieval-Augmented Generation). Next, the environment is set up, including entity relationship extraction, agent personality generation, and the configuration of simulation parameters. The simulation then runs in parallel, automatically interpreting prediction requests and dynamically updating the agents' memories over time. Finally, a "ReportAgent" generates a detailed prediction report, allowing for deep interaction with the simulated environment. Users can converse with individual agents or the ReportAgent to gain further insights.

The repository provides a clear vision for its application. It positions itself as a "pre-enactment laboratory" for decision-makers, allowing them to test policies and public relations strategies in a risk-free environment. For individual users, it serves as a creative sandbox, enabling them to explore different scenarios, such as alternative endings for novels or the exploration of complex ideas. The project aims to make prediction accessible and engaging, allowing users to visualize the potential outcomes of various "what-if" scenarios.

The README provides detailed instructions for both source code and Docker deployment. The source code deployment requires Node.js, Python, and the uv package manager. Users are instructed to configure environment variables, including API keys for LLMs (Large Language Models) and Zep Cloud, before installing dependencies and starting the services. Docker deployment offers a simplified approach, leveraging pre-built images and requiring only the configuration of environment variables. The project also provides links to demonstration videos showcasing its capabilities in predicting outcomes related to Wuhan University's public opinion and the potential lost ending of the classic novel "Dream of the Red Chamber." The project is actively seeking contributors and offers a QQ group for further communication and support. The project is supported by Shanda Group and acknowledges the contributions of the CAMEL-AI team for their OASIS framework.