MiroFish is a novel, multi-agent-based AI prediction engine designed to simulate and predict future outcomes by creating high-fidelity parallel digital worlds. It aims to overcome the limitations of traditional prediction methods by capturing emergent phenomena arising from individual interactions within a collective intelligence framework.

The core methodology of MiroFish involves a five-stage workflow:

1. Graph Construction (图谱构建):
   • Real-world Seed Extraction (现实种子提取): The process begins by extracting initial information, referred to as "seed information," from diverse real-world sources. This can range from structured data like news, policy drafts, and financial signals to unstructured narratives such as data analysis reports or fictional stories.
   • Individual and Group Memory Injection (个体与群体记忆注入): Relevant data is then processed and injected to form both individual agent memories and a collective group memory, which provides contextual understanding for the simulated environment.
   • GraphRAG Construction (GraphRAG构建): A knowledge graph is constructed, serving as a structured knowledge base for Retrieval Augmented Generation (RAG). This graph interlinks entities and their relationships, enabling agents to retrieve highly relevant and contextual information during their interactions.

2. Environment Setup (环境搭建):
   • Entity Relation Extraction (实体关系抽取): From the extracted seed data, entities and the relationships between them are identified and formalized, forming the foundational structure of the simulation environment.
   • Persona Generation (人设生成): Thousands of intelligent agents are generated, each imbued with unique personalities, long-term memory capabilities, and defined behavioral logics. These agents are the active participants in the simulated world.
   • Environment Configuration Agent Injection Simulation Parameters (环境配置Agent注入仿真参数): The simulated environment is configured, and the newly generated agents are injected into this environment along with specific simulation parameters that govern their initial states and interactions.

3. Start Simulation (开始模拟):
   • Dual-platform Parallel Simulation (双平台并行模拟): The simulation is executed, leveraging a multi-agent simulation engine driven by OASIS (with contributions acknowledged from CAMEL-AI). This stage involves parallel execution of agent behaviors and environmental dynamics.
   • Automatic Parsing of Prediction Requirements (自动解析预测需求): User-defined prediction requests, provided in natural language, are automatically parsed and translated into actionable objectives for the simulation.
   • Dynamic Update of Temporal Memory (动态更新时序记忆): Throughout the simulation, agents' individual and collective memories, along with the overall state of the digital world, are continuously and dynamically updated based on the unfolding events and interactions.

4. Report Generation (报告生成):
   • ReportAgent with Toolset for Interaction (ReportAgent拥有丰富的工具集与模拟后环境进行深度交互): A specialized ReportAgent is activated. This agent possesses a rich set of analytical tools, enabling it to deeply interact with the final state of the simulated environment and synthesize comprehensive findings.

5. Deep Interaction (深度互动):
   • Interaction with Simulated Agents (与模拟世界中的任意一位进行对话): Users can directly engage in dialogue with any individual agent within the simulated world, allowing for granular exploration of agent behaviors and perspectives.
   • Interaction with ReportAgent (与ReportAgent进行对话): Users can also interact with the ReportAgent to obtain detailed prediction reports, query specific aspects of the simulation, or gain summarized insights from the simulated outcomes.

MiroFish is designed for a wide range of applications, from serious predictive analysis (e.g., policy foresight, public opinion analysis, financial market trends) to creative simulations (e.g., exploring narrative endings for stories). It allows users to dynamically inject variables and observe outcomes, effectively functioning as a "digital sandbox" for "what-if" scenarios, making predictions and scenario planning accessible and interactive.