assafelovic/gpt-researcher
GPT Researcher ranked #1 in citation quality, report quality, and coverage in Carnegie Mellon University's DeepResearchGym benchmark — beating Perplexity, OpenAI's deep research, and HuggingFace with 85.36% citation precision and 90.82% recall. That alone should tell you this isn't another wrapper around ChatGPT. Here's the problem it solves: you ask an LLM a research question, and you get a confident-sounding answer built on stale training data, zero citations, and whatever hallucinations the model feels like inventing. GPT Researcher takes a different approach. It spawns autonomous planner and execution agents that fan out across 20+ web sources, scrape JavaScript-rendered pages, aggregate findings, and synthesize everything into a 2,000+ word report with inline citations. The architecture uses a tree-like exploration pattern — planners generate research sub-questions, executors gather evidence for each, and a publisher agent stitches the results into a coherent document. The LLM flexibility is where it gets interesting. You're not locked into OpenAI — it supports 100+ models including Anthropic Claude, Groq, and Llama 3. Search backends are equally flexible: Google, Bing, Tavily, or DuckDuckGo. Need to research internal documents alongside web sources? Local document search handles hybrid queries out of the box. With 25.8K GitHub stars, 3.4K forks, and nearly 3,000 commits on the main branch, this is one of the most actively maintained AI research agents in the open-source ecosystem. Meta, Google, Nvidia, Stanford, Harvard, Microsoft, MIT, and IBM all use it. The project ships with a NextJS frontend, Docker support, MCP server integration, multi-agent framework compatibility (LangGraph, AG2), and exports to PDF, Word, Markdown, JSON, and CSV. Deep Research mode runs recursive exploration that takes about 5 minutes per task at roughly $0.40 using o3-mini — a fraction of what manual research costs in developer time. Getting started requires Python 3.11+, an OpenAI key, and a Tavily API key. Clone, install dependencies, and you're running research queries in under 5 minutes.
Why It Matters
If you've tried building a research pipeline with raw LLM calls, you know the pain: hallucinated sources, shallow single-pass answers, and token limits that force you to chunk everything. GPT Researcher eliminates all three problems with a multi-agent architecture that actually verifies information across 20+ sources before writing a single paragraph. The CMU DeepResearchGym benchmark proved it — #1 in citation quality, report quality, and coverage against every major competitor including Perplexity and OpenAI's own deep research feature. For teams building knowledge-intensive applications — market analysis tools, competitive intelligence dashboards, due diligence automation — this is the most battle-tested open-source foundation available. The fact that it supports 100+ LLMs and multiple search backends means you avoid vendor lock-in while getting research quality that enterprise teams at Meta and Microsoft already trust in production.