deepseek

DeepSeek API Skill

Overview

DeepSeek provides state-of-the-art open reasoning and conversational language models via an OpenAI-compatible REST API. Key models include deepseek-chat (DeepSeek-V3) and deepseek-reasoner (DeepSeek-R1).

Installation

npm install openai
pip install openai

Authentication & Client Setup

import OpenAI from 'openai';

const deepseek = new OpenAI({
  apiKey: process.env.DEEPSEEK_API_KEY,
  baseURL: 'https://api.deepseek.com',
});

Core API Operations

1. General Conversational Chat (DeepSeek-V3)

const response = await deepseek.chat.completions.create({
  model: 'deepseek-chat',
  messages: [
    { role: 'system', content: 'You are an expert software architect.' },
    { role: 'user', content: 'Explain transaction isolation levels in PostgreSQL.' },
  ],
  temperature: 0.7,
  max_tokens: 2048,
});

console.log(response.choices[0].message.content);

2. Deep Reasoning Stream (DeepSeek-R1 with Chain-of-Thought)

const stream = await deepseek.chat.completions.create({
  model: 'deepseek-reasoner',
  messages: [
    { role: 'user', content: 'Prove that the square root of 2 is irrational.' },
  ],
  stream: true,
});

for await (const chunk of stream) {
  // Reasoning chain tokens
  const reasoning = (chunk.choices[0]?.delta as any)?.reasoning_content;
  if (reasoning) {
    process.stdout.write(reasoning);
  }
  // Final output content tokens
  const content = chunk.choices[0]?.delta?.content;
  if (content) {
    process.stdout.write(content);
  }
}

3. Structured JSON Output

const response = await deepseek.chat.completions.create({
  model: 'deepseek-chat',
  messages: [
    { role: 'system', content: 'Extract entities and output valid JSON.' },
    { role: 'user', content: 'John bought 3 apples from Whole Foods for $4.50' },
  ],
  response_format: { type: 'json_object' },
});

AI Pitfalls & Anti-Hallucination Guidelines

• Custom SDK Hallucination: There is no official @deepseek/sdk npm package. Always use the official openai package with baseURL: 'https://api.deepseek.com'.
• Reasoning Content Ingestion: On deepseek-reasoner, reasoning thoughts arrive under delta.reasoning_content, not delta.content.
• System Prompt Restrictions: deepseek-reasoner discourages heavy system prompt steering that constrains reasoning chains.