Deep Research.....but Open Source

OpenAI’s “Deep research” promises slower, more verifiable answers—often taking 5 to 30 minutes—by doing multi-step web dives with citations, rather than returning a quick, sometimes shaky summary. The catch is cost: using it requires ChatGPT Pro at $200 per month, which pushes many users to look for alternatives that deliver similar “research-grade” output without the subscription bill.

A workaround surfaced via an open-source implementation credited to David (named in the transcript as “David came up with his own open source implementation”). The pitch is straightforward: replicate Deep research’s workflow—planning a research task, pulling in sources from the web, and producing a markdown report with links—by running the system locally. Instead of paying a monthly premium for the feature, the setup relies on an OpenAI API key (pay-as-you-go) plus a web crawling service. In practice, the transcript frames this as “same capability” with a different cost structure: API usage can be pennies for light testing, while heavy usage can still get expensive.

The transcript also emphasizes what makes Deep research feel different from standard chat responses. Rather than “going off to the races,” it asks for additional context when needed, performs iterative searches, and can analyze multiple data types—text, images, and PDFs—described as multimodal. Citations are a core selling point: most claims come with links so users can verify the trail of evidence. The slower pace is treated as a feature, not a bug; the system’s multi-step reasoning and source gathering are presented as closer to how a human researcher would work.

To demonstrate the open-source version, the creator walks through setup on GitHub using Node.js inside a Docker container. The required components are a Fir Crawler API key (with a free hosted option up to 500 credits) and an OpenAI API key. After cloning the repo, creating environment variables, and building Docker configuration files (including Dockerfile and docker-compose.yml), the system is run from the command line. A sample prompt—“Which animal is better, cats or dogs?”—triggers a research plan with adjustable “breadth” and “depth” parameters, then outputs a markdown report.

The results are compared to ChatGPT Deep research: both conclude “dogs are better,” with the open-source output appearing more command-line and link-heavy, while ChatGPT offers a more polished interface. A second test asks which operating system is better for IT professionals—Mac, Linux, or Windows—framing the criteria around cost, licensing, and security. The open-source system lands on a hybrid strategy: use each platform’s strengths while applying platform-agnostic security practices.

Finally, the transcript notes a broader trend: Hugging Face is reportedly working on an agentic framework for deep search to compete with OpenAI’s approach, underscoring how quickly the ecosystem is moving toward citation-backed, tool-using research agents.