Byte's "Button Space mini version" is fully open source, has super deep research capabilities, and also integrates voice podcast generation functions!
ByteDance open-sources DeerFlow, a multi-agent framework that combines deep research and voice podcasts, and is a new tool for exploring complex problems.
Core content:
1. The community-driven nature of the DeerFlow framework and its GitHub open source status
2. The integrated function of deep research, podcast generation, and PPT production
3. Analysis of the technical architecture and core modules, as well as a quick start guide
Another major open source project is coming!
DeerFlow (Deep Exploration and Efficient Research Flow) launched by ByteDance is now fully open source on GitHub.
DeerFlow is a community-driven deep learning framework, an open-source replica of ButtonSpace, and also integrates the same speech synthesis capabilities as Doubao.
The project has received 2100+ Stars and received rave reviews from the open source community!
Like Buttonspace, DeepFlow is a multi-agent framework that combines large language models with specialized tools such as web search, crawling, and Python code execution.
The following features are included:
In-depth research : Conduct comprehensive online research on any topic
Podcast generation : Automatically generate podcast scripts and synthesize speech based on research reports
PPT Maker : Automatically create simple PowerPoint presentations
Code Analysis : Execute and analyze code using the Python REPL tool
Just look at the official demonstration video and you will understand that DeerFlow is not a simple web search, but an in-depth exploration of complex problems.
Enter a topic and it will search the web, aggregate information, write a report, and even turn the report into an audio podcast similar to Google Notebookllm.
The technical architecture is as follows:
DeepFlow builds a flexible state-based workflow framework based on LangGraph.
The system consists of the following core modules:
Coordinator : The entry point of the entire system, managing the workflow lifecycle
Planner : Strategic component responsible for task decomposition and planning
Research Team : It consists of three parts, each with its own responsibilities, including researchers, programmers and reporters.
The researcher role uses tools such as web searches and crawlers to collect information, which is then handed over to the programmer role to use Python to handle code analysis and technical tasks. Finally, the summarized information is handed over to the reporter to generate relevant reports.
Audio Podcast
DeerFlow now integrates the same text-to-speech (TTS) function as Doubao, which can convert research reports into high-quality speech.
Using the volcengine TTS API, the speed, volume, and pitch can all be customized, turning text into podcast content with one click!
Ready to use
It only takes a few steps to get started with DeerFlow:
# Clone the repository git clone https://github.com/bytedance/deer-flow.git cd deer-flow # Install dependencies uv sync # Configure API key cp .env.example .env # Configure LLM model cp conf.yaml.example conf.yaml # Run the project uv run main.py
Support multiple search engines:
- Tavily (default)
- DuckDuckGo
- Brave Search
- Arxiv (academic paper search)
