groq-api

Groq API

使用 Groq 的超高速大模型推理（300-1000+ tokens/sec）构建应用。

快速入门

安装

# Python
pip install groq

# TypeScript/JavaScript
npm install groq-sdk

环境设置

export GROQ_API_KEY=<your-api-key>

基本聊天补全

Python：

from groq import Groq

client = Groq()  # Uses GROQ_API_KEY env var

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hello"}]
)
print(response.choices[0].message.content)

TypeScript：

import Groq from "groq-sdk";

const client = new Groq();

const response = await client.chat.completions.create({
    model: "llama-3.3-70b-versatile",
    messages: [{ role: "user", content: "Hello" }],
});
console.log(response.choices[0].message.content);

模型选择

用例	模型	备注
快速且便宜	llama-3.1-8b-instant	最适合简单任务
均衡	llama-3.3-70b-versatile	质量与成本平衡
最高质量	openai/gpt-oss-120b	内置工具 + 推理
Agentic	groq/compound	网页搜索 + 代码执行
推理	openai/gpt-oss-20b	快速推理（low/med/high）
视觉/OCR	llama-4-scout-17b-16e-instruct	图像理解
音频 STT	whisper-large-v3-turbo	转录
TTS	playai-tts	文本转语音

完整模型列表和定价请参阅 references/models.md。

常见模式

流式响应

stream = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Tell me a story"}],
    stream=True
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

系统消息

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello"}
    ]
)

异步客户端（Python）

import asyncio
from groq import AsyncGroq

async def main():
    client = AsyncGroq()
    response = await client.chat.completions.create(
        model="llama-3.3-70b-versatile",
        messages=[{"role": "user", "content": "Hello"}]
    )
    return response.choices[0].message.content

print(asyncio.run(main()))

JSON 模式

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "List 3 colors as JSON array"}],
    response_format={"type": "json_object"}
)

结构化输出（JSON Schema）

强制输出符合某个架构。提供两种模式：

模式	保证	模型
strict: true	100% 符合架构	openai/gpt-oss-20b、openai/gpt-oss-120b
strict: false	尽力符合架构	所有受支持的模型

严格模式（保证符合）：

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "Extract: John is 30 years old"}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "person",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "name": {"type": "string"},
                    "age": {"type": "integer"}
                },
                "required": ["name", "age"],
                "additionalProperties": False
            }
        }
    }
)

使用 Pydantic：

from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "Extract: John is 30"}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "person",
            "strict": True,
            "schema": Person.model_json_schema()
        }
    }
)
person = Person.model_validate(json.loads(response.choices[0].message.content))

有关架构要求、验证库和示例，请参阅 references/structured-outputs.md。

音频

转录（语音转文本）

with open("audio.mp3", "rb") as f:
    transcription = client.audio.transcriptions.create(
        model="whisper-large-v3-turbo",
        file=f,
        language="en",  # Optional: ISO-639-1 code
        response_format="verbose_json",  # json, text, verbose_json
        timestamp_granularities=["word", "segment"]
    )
print(transcription.text)

翻译（为英语）

with open("french_audio.mp3", "rb") as f:
    translation = client.audio.translations.create(
        model="whisper-large-v3",
        file=f
    )
print(translation.text)  # English text

文本转语音

response = client.audio.speech.create(
    model="playai-tts",
    input="Hello, world!",
    voice="Fritz-PlayAI",
    response_format="wav",  # flac, mp3, mulaw, ogg, wav
    speed=1.0  # 0.5 to 5
)
response.write_to_file("output.wav")

视觉

使用 Llama 4 多模态模型处理图像。每个请求最多支持 5 张图像。

模型： meta-llama/llama-4-scout-17b-16e-instruct（更快）、meta-llama/llama-4-maverick-17b-128e-instruct（质量更高）

URL 中的图像

response = client.chat.completions.create(
    model="meta-llama/llama-4-scout-17b-16e-instruct",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What's in this image?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}}
        ]
    }]
)

本地图像（Base64）

import base64

def encode_image(path: str) -> str:
    with open(path, "rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")

response = client.chat.completions.create(
    model="meta-llama/llama-4-scout-17b-16e-instruct",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Describe this image"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encode_image('photo.jpg')}"}}
        ]
    }]
)

OCR / 将数据提取为 JSON

response = client.chat.completions.create(
    model="meta-llama/llama-4-scout-17b-16e-instruct",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Extract all text and data as JSON"},
            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{base64_image}"}}
        ]
    }],
    response_format={"type": "json_object"}
)

有关多图像、图像工具使用和多轮对话，请参阅 references/vision.md。

工具使用

有关工具调用模式和示例，请参阅 references/tool-use.md。

快速示例：

import json

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get weather for a location",
        "parameters": {
            "type": "object",
            "properties": {"location": {"type": "string"}},
            "required": ["location"]
        }
    }
}]

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Weather in Paris?"}],
    tools=tools
)

if response.choices[0].message.tool_calls:
    for tc in response.choices[0].message.tool_calls:
        args = json.loads(tc.function.arguments)
        # Execute function and continue conversation

内置工具（Agentic）

使用 groq/compound 或 openai/gpt-oss-120b 执行内置网页搜索和代码：

response = client.chat.completions.create(
    model="groq/compound",
    messages=[{"role": "user", "content": "Search for latest Python news"}]
)
# Model automatically uses web search

MCP（远程工具）

连接第三方 MCP 服务，以使用 Stripe、GitHub、网页抓取等工具。请使用 Responses API：

import openai

client = openai.OpenAI(
    api_key=os.environ.get("GROQ_API_KEY"),
    base_url="https://api.groq.com/openai/v1"
)

response = client.responses.create(
    model="openai/gpt-oss-120b",
    input="What models are trending on Huggingface?",
    tools=[{
        "type": "mcp",
        "server_label": "Huggingface",
        "server_url": "https://huggingface.co/mcp"
    }]
)

有关 MCP 配置和常用服务，请参阅 references/tool-use.md。

推理模型

控制模型如何思考复杂问题。

模型： openai/gpt-oss-20b、openai/gpt-oss-120b（low/medium/high）、qwen/qwen3-32b（none/default）

带推理力度设置的 GPT-OSS

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "How many r's in strawberry?"}],
    reasoning_effort="high",  # low, medium, high
    temperature=0.6,
    max_completion_tokens=1024
)

print(response.choices[0].message.content)
print("Reasoning:", response.choices[0].message.reasoning)

带解析推理的 Qwen3

response = client.chat.completions.create(
    model="qwen/qwen3-32b",
    messages=[{"role": "user", "content": "Solve: x + 5 = 12"}],
    reasoning_format="parsed"  # raw, parsed, hidden
)

print("Answer:", response.choices[0].message.content)
print("Reasoning:", response.choices[0].message.reasoning)

隐藏推理（GPT-OSS）

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "What is 15% of 80?"}],
    include_reasoning=False  # Hide reasoning in response
)

有关流式处理、推理时的工具使用和最佳实践，请参阅 references/reasoning.md。

批处理

用于高容量异步处理（完成时间窗口为 24h-7d）：

# 1. Create JSONL file with requests
# 2. Upload file
# 3. Create batch
batch = client.batches.create(
    input_file_id=file_id,
    endpoint="/v1/chat/completions",
    completion_window="24h"
)

# 4. Check status
batch = client.batches.retrieve(batch.id)
if batch.status == "completed":
    results = client.files.content(batch.output_file_id)

有关批处理 API 的完整详情，请参阅 references/api-reference.md。

提示缓存

自动将重复提示前缀的延迟和成本降低 50%。无需更改代码。

支持的模型： moonshotai/kimi-k2-instruct-0905、openai/gpt-oss-20b、openai/gpt-oss-120b、openai/gpt-oss-safeguard-20b

工作原理：

• 将静态内容（系统提示词、工具、示例）放在开头
• 将动态内容（用户查询）放在末尾
• 缓存会自动匹配前缀并应用 50% 折扣
• 缓存闲置 2 小时后过期

跟踪缓存用量：

response = client.chat.completions.create(
    model="moonshotai/kimi-k2-instruct-0905",
    messages=[{"role": "system", "content": large_system_prompt}, ...]
)

cached = response.usage.prompt_tokens_details.cached_tokens
print(f"Cached tokens: {cached}")  # 50% discount applied to these

有关优化策略和示例，请参阅 references/prompt-caching.md。

内容审核

使用安全防护模型检测并过滤有害内容。

Llama Guard 4

通用内容安全分类。返回 safe 或 unsafe\nSX（类别代码）。

response = client.chat.completions.create(
    model="meta-llama/Llama-Guard-4-12B",
    messages=[{"role": "user", "content": user_input}]
)

if response.choices[0].message.content.startswith("unsafe"):
    # Block or handle unsafe content
    pass

GPT-OSS Safeguard 20B

使用自定义策略检测提示词注入。返回结构化 JSON。

response = client.chat.completions.create(
    model="openai/gpt-oss-safeguard-20b",
    messages=[
        {"role": "system", "content": injection_detection_policy},
        {"role": "user", "content": user_input}
    ]
)
# Returns: {"violation": 1, "category": "Direct Override", "rationale": "..."}

完整策略、有害内容分类体系和集成模式请参阅 references/moderation.md。

错误处理

from groq import Groq, RateLimitError, APIConnectionError, APIStatusError

client = Groq()

try:
    response = client.chat.completions.create(
        model="llama-3.3-70b-versatile",
        messages=[{"role": "user", "content": "Hello"}]
    )
except RateLimitError:
    # Wait and retry with exponential backoff
    pass
except APIConnectionError:
    # Network issue
    pass
except APIStatusError as e:
    # API error (check e.status_code)
    pass

完整的音频 API 参考（包括文件处理、元数据字段和提示词编写指南）请参阅 references/audio.md。

资源

• 模型与定价：references/models.md
• 工具使用指南：references/tool-use.md
• 视觉指南：references/vision.md
• 音频指南：references/audio.md
• 推理指南：references/reasoning.md
• 结构化输出：references/structured-outputs.md
• 提示词缓存：references/prompt-caching.md
• 内容审核指南：references/moderation.md
• SDK 参考：references/sdk.md
• 完整 API 参考：references/api-reference.md
• 官方文档：https://console.groq.com/docs
• Python SDK：https://github.com/groq/groq-python
• TypeScript SDK：https://github.com/groq/groq-typescript