Deep Research clone that skips the fine-tuned o3

What it does

This is a Next.js chatbot template that clones OpenAI’s Deep Research experience. It queries the live web through Firecrawl, extracts structured data from multiple sites, and feeds the results to a configurable reasoning model—defaulting to o1-mini—to produce structured research analysis.

The interesting bit

Rather than training a bespoke model like o3, the project replicates the capability with clever plumbing: Firecrawl handles the messy work of search and extraction, while a general-purpose reasoning LLM does the analysis. The value is in the orchestration, not a new foundation model.

Key highlights

• Uses Firecrawl’s search and extract APIs to pull real-time, structured web data into the context window
• Built on the Vercel AI SDK, so swapping between OpenAI, Anthropic, Cohere, or OpenRouter is nominally a few lines of code
• Splits general chat from reasoning tasks via a separate REASONING_MODEL environment variable
• Bundles NextAuth, Vercel Postgres, and Vercel Blob for chat history, user data, and file storage
• Requires a stack of external services and API keys to function

Caveats

• Running it demands a laundry list of third-party accounts and API keys: Firecrawl, an LLM provider, Vercel Postgres, Vercel Blob, and more
• Vercel’s Hobby tier caps functions at 60 seconds, well below the default 300-second research timeout
• Using non-OpenAI reasoning models like DeepSeek-R1 forces you to disable JSON schema validation, which the README admits can yield less structured output

Verdict

Worth exploring if you want a hackable starting point for automated web research. Look elsewhere if you need a lightweight, self-contained agent.