MindSearch — Multi-Agent AI Search Engine

Introduction

MindSearch is an open-source AI-powered search engine developed by the InternLM team at Shanghai AI Lab. It uses a multi-agent system to break down complex questions into sub-queries, search the web in parallel, and produce well-sourced, comprehensive answers similar to how a human researcher would work.

What MindSearch Does

• Decomposes complex user questions into a directed acyclic graph (DAG) of sub-queries
• Dispatches multiple search agents in parallel to gather information for each sub-query
• Aggregates and synthesizes results from dozens of web pages into a coherent answer
• Provides source citations and a visual graph of the reasoning process
• Supports multiple LLM backends and search engines including SearXNG and Bing

Architecture Overview

MindSearch consists of two core agent roles: a Planner and multiple Searchers. The Planner receives the user query and generates a DAG of atomic sub-questions with dependency edges. Independent sub-questions are dispatched to Searcher agents in parallel. Each Searcher calls a web search API, reads the returned pages, and extracts relevant information. Once all sub-tasks complete, the Planner synthesizes the individual findings into a final answer with citations. The system is built on the InternLM agent framework and uses a React-based frontend for the visual interface.

Self-Hosting & Configuration

• Requires Python 3.9+ and a running SearXNG instance or Bing API key for web search
• Configure the LLM backend via environment variables (supports InternLM, OpenAI, and compatible APIs)
• Deploy SearXNG locally with Docker for a fully self-hosted, privacy-respecting setup
• Adjust the maximum number of parallel searcher agents and search depth in the config
• The frontend can be run separately with Node.js or served through the included Gradio/Streamlit interface

Key Features

• Graph-based query decomposition that handles multi-hop, comparative, and aggregation questions
• Parallel search execution dramatically reduces latency compared to sequential approaches
• Visual reasoning graph showing how the answer was assembled from sub-queries
• Model-agnostic design works with both commercial and open-weight LLMs
• Self-hostable with local search via SearXNG for complete data privacy

Comparison with Similar Tools

• Perplexica — single-agent search-and-answer; MindSearch uses multi-agent parallel decomposition for complex queries
• SearchGPT — closed commercial product; MindSearch is fully open source and self-hostable
• SearXNG — privacy metasearch engine without AI synthesis; MindSearch adds LLM-based reasoning on top
• Tavily — search API for AI agents; MindSearch is a full search application, not just an API
• Stanford STORM — research report generation; MindSearch focuses on real-time interactive question answering

FAQ

Q: Can I use MindSearch without an internet connection? A: No. MindSearch requires live web search to retrieve current information. However, the search and LLM components can all be self-hosted on your network.

Q: Which search backends are supported? A: SearXNG (recommended for self-hosting), Bing Web Search API, and DuckDuckGo.

Q: How many web pages does it process per query? A: Depending on query complexity, MindSearch typically reads and synthesizes content from 20 to 60 web pages per question.

Q: Can I use GPT-4 or Claude as the reasoning model? A: Yes. Any OpenAI-compatible API can be configured as the LLM backend.

Sources

• https://github.com/InternLM/MindSearch
• https://github.com/InternLM/MindSearch/blob/main/README.md