assafelovic/gpt-researcher

What it solves

Manual research is often time-consuming and resource-intensive. Standard LLMs frequently suffer from hallucinations due to outdated training data, token limits that prevent long-form reporting, and a reliance on limited web sources that can introduce bias or shallow results.

How it works

GPT Researcher uses a multi-agent architecture consisting of 'planner' and 'execution' agents. The planner breaks a research query into specific research questions, and execution agents crawl the web or local documents to gather information. A publisher agent then aggregates these findings into a comprehensive, cited report.

To ensure quality, it employs "Deep Research" for recursive exploration of subtopics and a smart context filtering system (using Jev or keyword ranking) to discard irrelevant scraped content before it reaches the LLM. It can also integrate with the Model Context Protocol (MCP) to access specialized data sources like GitHub or databases.

Who it’s for

Individuals and organizations needing detailed, factual, and unbiased research reports on complex topics without spending weeks on manual data gathering.

Highlights

• Deep Research: A recursive, tree-like exploration pattern that dives into subtopics for greater depth and breadth.
• Comprehensive Reporting: Generates reports exceeding 2,000 words with citations from over 20 sources.
• Hybrid Data Sources: Supports both web scraping and local document research (PDF, Word, Excel, etc.).
• Smart Filtering: Uses Jev to score passage relevance, reducing noise and cost compared to standard embeddings.
• AI Visuals: Automatically generates and embeds professional inline illustrations using Google Gemini.
• Multi-Agent Frameworks: Built-in support for LangGraph and AG2 to coordinate specialized agents.