Groq Core Workflow A: Chat, Tools & Structured Output

Overview

Primary integration patterns for Groq: chat completions, tool/function calling, JSON mode, and structured outputs. Groq's LPU delivers sub-200ms time-to-first-token, making these patterns viable for real-time user-facing features. This skill walks through five workflow steps; the lean skeleton lives here, and the full copy-paste code lives in references/.

Prerequisites

• Install the SDK with npm install groq-sdk.
• Set GROQAPIKEY in the environment (see Authentication below).
• Familiarity with the Groq model line-up and which model fits each task.

Authentication

Groq authenticates via an API key. Create one at console.groq.com/keys and

export it as GROQAPIKEY; the SDK reads it automatically, so new Groq()

needs no explicit argument. Never hardcode the key — read it from the

environment (or a secrets manager) so it stays out of source control.

Model Selection for This Workflow

Instructions

Work through the five patterns in order. Read the target file, then Write or

Edit the integration code into your project.

1. Chat completion — send system + user messages to

groq.chat.completions.create and return choices[0].message.content plus

usage. Skeleton below; full example in

worked examples.

1. Tool use / function calling — a three-phase loop: send the message with

tools + toolchoice: "auto", execute any returned toolcalls, then send

the results back for the final answer. Full code in

implementation.

1. JSON mode — set responseformat: { type: "jsonobject" } and describe

the JSON shape in the system prompt. See

implementation.

1. Structured outputs — use responseformat.jsonschema with

strict: true for guaranteed schema compliance (no post-validation). See

implementation.

1. Multi-turn conversation — accumulate the message history and push each

assistant reply back onto the stack. See

worked examples.

Minimal chat skeleton:

import Groq from "groq-sdk";
const groq = new Groq();

const completion = await groq.chat.completions.create({
  model: "llama-3.3-70b-versatile",
  messages: [
    { role: "system", content: "You are a concise technical assistant." },
    { role: "user", content: userMessage },
  ],
  temperature: 0.7,
  max_tokens: 1024,
});
// completion.choices[0].message.content, completion.usage

Output

Each pattern returns a predictable shape:

• Chat completion — { reply: string, usage: {...} }; usage carries

prompttokens / completiontokens for cost metering.

• Tool use — the final assistant content string, produced after the tool

results are fed back; intermediate tool_calls carry function.name and a

JSON-string function.arguments.

• JSON mode — a parsed JavaScript object matching the shape described in the

system prompt (parse message.content with JSON.parse).

• Structured outputs — a parsed object guaranteed to satisfy the declared

JSON schema, so no downstream validation is required.

• Multi-turn — the latest reply string, with conversation state retained in

the class instance for the next turn.

Error Handling

Examples

The chat skeleton above is the smallest complete call. Two fuller runnable

examples live in worked examples:

• Example 1 — Chat completion with system prompt + rolling history, returning

reply and token usage.

• Example 2 — Multi-turn conversation class that retains context across turns.

For tool use, JSON mode, and strict structured outputs, see

full implementation.

Resources

• Groq Tool Use Docs
• Groq Structured Outputs
• Groq Text Generation
• Full implementation — tool use, JSON mode, structured outputs
• Worked examples — chat completion, multi-turn conversation

Next Steps

For audio, vision, and speech workflows, see the companion groq-core-workflow-b

skill, which covers Whisper transcription, vision inputs, and text-to-speech.