Build your own Perplexity with LLM agents that search in parallel

What it does

MindSearch is a multi-agent web search framework built on the Lagent library. It takes a user query, deploys concurrent LLM agents to search across engines like DuckDuckGo, Bing, Brave, Google, or Tencent, and returns an answer. The project ships with a FastAPI backend and three frontend options—React, Gradio, and Streamlit—plus a terminal debug mode for headless use.

The interesting bit

Instead of a single monolithic LLM call, the framework distributes work across specialized agents. The recent refactor to Lagent v0.5 specifically targeted better concurrency, suggesting the authors treat parallel multi-query search as a first-class feature rather than an afterthought.

Key highlights

• Supports five search engines out of the box, though most require API keys (DuckDuckGo is the exception).
• Backend runs on FastAPI with asynchronous agents; frontends include React, Gradio, and Streamlit.
• Works with local models (InternLM2.5-7b-chat) or commercial APIs like GPT-4; extending to other models requires editing models.py.
• Includes a live demo and an arXiv paper describing the architecture.
• Apache 2.0 licensed.

Caveats

• Adding non-standard models means editing Python source rather than dropping in a config file.
• The React frontend needs manual backend URL injection via sed before it will compile.
• Search engine credentials are scattered across multiple environment variable names depending on the provider.

Verdict

Good fit for developers who want a self-hosted Perplexity-style search with transparent agent wiring. Skip it if you need a plug-and-play SaaS; the README is honest about the assembly required.