Open Deep Research

An Open-Source clone of Open AI's Deep Research experiment. Instead of using a fine-tuned version of o3, this method uses Firecrawl's extract + search with a reasoning model to deep research the web.

Check out the demo here

Features

• Firecrawl Search + Extract
• Feed realtime data to the AI via search
• Extract structured data from multiple websites via extract
• Next.js App Router
• Advanced routing for seamless navigation and performance
• React Server Components (RSCs) and Server Actions for server-side rendering and increased performance
• AI SDK
• Unified API for generating text, structured objects, and tool calls with LLMs
• Hooks for building dynamic chat and generative user interfaces
• Supports OpenAI (default), Anthropic, Cohere, and other model providers
• shadcn/ui
• Styling with Tailwind CSS
• Component primitives from Radix UI for accessibility and flexibility
• Data Persistence
• Vercel Postgres powered by Neon for saving chat history and user data
• Vercel Blob for efficient file storage
• NextAuth.js
• Simple and secure authentication

Model Providers

This template ships with OpenAI gpt-4o as the default. However, with the AI SDK, you can switch LLM providers to OpenAI, Anthropic, Cohere, and many more with just a few lines of code.

This repo is compatible with OpenRouter and OpenAI. To use OpenRouter, you need to set the OPENROUTER_API_KEY environment variable.

Function Max Duration

By default, the function timeout is set to 300 seconds (5 minutes). If you're using Vercel's Hobby tier, you'll need to reduce this to 60 seconds. You can adjust this by changing the MAX_DURATION environment variable in your .env file:

MAX_DURATION=60

Learn more about it here

Deploy Your Own

You can deploy your own version of the Next.js AI Chatbot to Vercel with one click:

Running locally

You will need to use the environment variables defined in .env.example to run Next.js AI Chatbot. It's recommended you use Vercel Environment Variables for this, but a .env file is all that is necessary.

Note: You should not commit your .env file or it will expose secrets that will allow others to control access to your various OpenAI and authentication provider accounts.

1. Install Vercel CLI: npm i -g vercel
2. Link local instance with Vercel and GitHub accounts (creates .vercel directory): vercel link
3. Download your environment variables: vercel env pull

1. First install all dependencies

pnpm install

2. Then run database migrations

pnpm db:migrate

3. Run the app

pnpm dev

Your app template should now be running on localhost:3000.

Models dependencies

If you want to use a model other than the default, you will need to install the dependencies for that model.

TogetherAI's Deepseek:

pnpm add @ai-sdk/togetherai

Note: Maximum rate limit https://docs.together.ai/docs/rate-limits

Reasoning Model Configuration

The application uses a separate model for reasoning tasks (like research analysis and structured outputs). This can be configured using the REASONING_MODEL environment variable.

Available Options

Important Notes

• Only certain OpenAI models (gpt-4o, o1, o3-mini) natively support structured JSON outputs
• Other models (deepseek-reasoner) can be used but may require disabling JSON schema validation
• When using models that don't support JSON schema:
• Set BYPASS_JSON_VALIDATION=true in your .env file
• This allows non-OpenAI models to be used for reasoning tasks
• Note: Without JSON validation, the model responses may be less structured
• The reasoning model is used for tasks that require structured thinking and analysis, such as:
• Research analysis
• Document suggestions
• Data extraction
• Structured responses
• If no REASONING_MODEL is specified, it defaults to o1-mini
• If an invalid model is specified, it will fall back to o1-mini

Usage

Add to your .env file:

# Choose one of: deepseek-reasoner, deepseek-ai/DeepSeek-R1
REASONING_MODEL=deepseek-ai/DeepSeek-R1

# Required when using models that don't support JSON schema (like deepseek-reasoner)
BYPASS_JSON_VALIDATION=true

The reasoning model is automatically used when the application needs structured outputs or complex analysis, regardless of which model the user has selected for general chat.