DeepSeek Reasoning + OpenAI summarization

r1-openai.mp4

As of 1/26/2025, DeepSeek's R1 model does not support structured outputs and tool calling.. This project is a proof of concept for how to use DeepSeek's reasoning model to analyze questions in detail and then use GPT's summarization power to create a structured output and make tool calls while keeping the cost down.

Project Description

A Next.js application that combines DeepSeek's reasoning capabilities with GPT's summarization power. This project demonstrates how to:

1. Use DeepSeek's reasoning stage and handoff to OpenAI
2. Process responses in multiple formats (basic, structured, tool-calling)
3. Swap GPT models on the client to experiement with summarization and cost

Features

• Modern Next.js web interface with real-time interactions
• Multiple API endpoints for different response formats
• Dark/light theme support with Tailwind CSS
• Real-time response streaming
• Structured data outputs with JSON schema validation
• Cost-effective hybrid model approach
• Chat UI made using shadcn/ui

Development

pnpm install 
touch .env
pnpm dev

Environment variables:

DEEPSEEK_API_KEY=
OPENAI_API_KEY=

API Routes

Basic API (/api/route.ts)

• Simple question/answer format
• Uses DeepSeek for reasoning and GPT-3.5-turbo for summarization
• Returns plain text responses with token usage metrics

Structured API (/api/structured/route.ts)

• Enhanced response format with JSON schema validation
• Structured output includes:
  • One-sentence summary
  • Bullet points
  • Reasoning steps count
  • Follow-up prompts
• Uses GPT-4o-mini for structured summarization

Tool Calling API (/api/tool-calling/route.ts)

• Advanced response format with function calling capabilities
• Similar structure to structured API
• Optimized for integration with external tools and services

Requirements

You'll need API keys for:

• DeepSeek API (for reasoning capabilities)
• OpenAI API (for summarization)

How It Works

1. User submits a question through the web interface or API
2. The selected API route processes the request:
   • DeepSeek model provides detailed reasoning
   • GPT model creates summary after thinking phase and returns API response
3. Usage metrics and timing are tracked and returned as well

Implementation Details

The code demonstrates several key patterns:

• Next.js App Router architecture
• React Server Components with client-side interactivity
• Tailwind CSS for styling with CSS variables
• Radix UI primitives for accessible components
• TypeScript for type safety
• Real-time response streaming
• Error handling and loading states