InternLM/MindSearch
MindSearch
MindSearch is an LLM-based multi-agent search engine that decomposes complex user questions into targeted sub-queries and routes each to a specialized agent for parallel investigation. The system orchestrates multiple agents through a large language model, coordinating their tasks and interpreting search results to produce coherent answers from multiple sources.
The project provides a configurable search backend interface that allows switching between Google, DuckDuckGo, Brave, and Bing search APIs by updating a configuration attribute. It includes a terminal-based debug interface for testing and troubleshooting the multi-agent search pipeline during development, enabling headless backend execution without a graphical frontend.
The system synthesizes findings from parallel sub-query agents into a single coherent answer, addressing the original multi-faceted question. The documentation covers running the search backend without a frontend and debugging the search pipeline locally through the command-line interface.
Features
- Multi-Agent Search Engines - Runs a search engine powered by large language models that decomposes queries and synthesizes results from multiple sources.
- LLM-Driven Agent Loops - Uses a large language model to coordinate agent tasks, route sub-queries, and interpret search results.
- Multi-Agent Search Engines - An AI search engine that decomposes complex queries into sub-questions and routes them to specialized agents for parallel investigation and answer synthesis.
- Multi-Agent Synthesis Systems - Collects findings from parallel sub-query agents and merges them into a single coherent answer.
- Multi-Agent Frameworks - Orchestrates multiple LLM agents to perform web searches, synthesize results, and produce coherent answers to multi-faceted questions.
- Question Answering - Collects and merges findings from multiple sub-queries into a single coherent answer addressing the original complex question.
- Recursive Query Decomposition - Splits complex user questions into parallel sub-queries handled by specialized agents.
- Sub-Query Decomposers - Splits multi-faceted questions into smaller, targeted sub-queries and routes each to a specialized agent for parallel investigation.
- Multi-Query Answer Synthesizers - Synthesizes findings from parallel sub-query agents into a single coherent answer for multi-faceted questions.
- Search Backend Swapping - Changes the web search API used for retrieving results by updating a configuration attribute, supporting Google, DuckDuckGo, Brave, and Bing.
- Web Search API Swappers - Switches between Google, DuckDuckGo, Brave, and Bing search APIs by updating a configuration attribute.
- AI Search Engines - Multi-agent framework for web search and reasoning.
- Search Engines - Multi-agent framework for web search and reasoning.
