Groq integration

AI: ultra-low-latency LLM inference (Llama, Mixtral, Gemma); OpenAI-compatible API

Help CentreConnectors

Overview

Groq runs open-weight LLMs (Llama, Mixtral, Whisper) on LPU hardware that delivers very low latency. The Automize connector covers Chat Completions, transcription, translation, and model metadata via Groq's OpenAI-compatible API.

  • Regions: Single global API at api.groq.com. The OpenAI-compatible surface lives at /openai/v1/* – same request shape as OpenAI Chat.
  • Plans: Free tier has generous per-model RPM/TPM but caps daily tokens. Paid tier (on-demand) removes the daily cap and raises RPM. Whisper STT (transcribe + translate) bills per audio second.

Setting up the connection

Auth type: Bearer Token.

  1. In Groq Console → API Keys, click Create API Key (starts with gsk_).
  2. Paste as the API token. Connector sends Authorization: Bearer <key>.

Find your credentials at https://console.groq.com/keys.


Rate limits

Per-model RPM (requests/min), TPM (tokens/min), RPD (requests/day), TPD (tokens/day) – all four are tracked separately and any can fire 429. The X-RateLimit-Remaining-* headers expose live state; connector backs off on 429.


Data model

Groq's surface is intentionally narrow – it's an inference layer,
not a model owner. You pick a model + send a request:

    Chat                OpenAI-compatible chat completions
    Audio Transcribe    Whisper STT, source language preserved
    Audio Translate     Whisper-to-English (target=English only)
    Models              metadata + capability flags

Models accessible (subject to availability changes):

    llama-3.3-70b-versatile   general-purpose, tool-use capable
    llama-3.1-8b-instant      tiny + fast, weaker reasoning
    llama-3.2-* (vision)      multimodal variants
    mixtral-8x7b-32768        long context, sparse-MoE
    whisper-large-v3          STT (transcribe)
    whisper-large-v3-turbo    STT, faster, lossy on rare languages

Tool/function calling support is per-model. llama-3.3-70b-versatile
is the most reliable; smaller models occasionally hallucinate tool
names or skip required arguments.

Output streaming exists in the API; this connector resolves the
full response per call rather than streaming token-by-token.

Available operations

5 operations available. Click any row to jump to its detail.

OperationCategoryWhat it does
Chat CompletionGeneralGenerate a chat completion
Get ModelGeneralGet model details
List ModelsGeneralList available models
Transcribe AudioGeneralTranscribe audio file via Whisper
Translate AudioGeneralTranslate audio to English via Whisper

Operations

Chat Completion

Generate a chat completion

ParameterRequiredTypeDescription
Model modelNotext
Messages messagesNotext
Temperature temperatureNotext
Max Tokens max_tokensNotext
Top P top_pNotext
Stop stopNotext
Tools toolsNotext

Tips

  • OpenAI-compatible – same messages/role/content shape as OpenAI Chat. Tools/function calling supported on llama-3.1-* and llama-3.3-70b-versatile.
  • max_tokens caps output. temperature=0 for deterministic; default 1.0 can vary across identical inputs.
  • Stream support exists in the API but the connector resolves the full response.

Get Model

Get model details

ParameterRequiredTypeDescription
Model Id model_idYestext

Tips

  • GET /openai/v1/models/{model_id} – single model metadata.
  • Useful as a credential health-check (cheap, deterministic) and to confirm a model_id is still available before kicking off a chain of Chat calls.

List Models

List available models

No input parameters.

Tips

  • GET /openai/v1/models – returns every model accessible to the API key, with object='model', owned_by='Groq', and active flag.
  • context_window varies by model – pin to the listed max in production; over-context returns a generic 400.

Transcribe Audio

Transcribe audio file via Whisper

ParameterRequiredTypeDescription
File fileNotext
Model modelNotext
Language languageNotext
Prompt promptNotext
Response Format response_formatNotext
Temperature temperatureNotext

Tips

  • POST /openai/v1/audio/transcriptions – model=whisper-large-v3 or whisper-large-v3-turbo. Audio file (up to 25 MB).
  • response_format=verbose_json returns word-level timestamps; default is text.

Translate Audio

Translate audio to English via Whisper

ParameterRequiredTypeDescription
File fileNotext
Model modelNotext
Prompt promptNotext
Response Format response_formatNotext
Temperature temperatureNotext

Tips

  • Same as transcribe but outputs English regardless of source language. Use the English-target flow only – for other targets, transcribe then translate downstream.

FAQ

Why is Groq so much faster than OpenAI for the same model?
Different hardware. Groq's LPU is designed for sequential token generation – they trade batching efficiency for time-per-token. Expect 5-10x faster output for streaming workloads.
Are responses identical to OpenAI's hosted Llama?
Same weights, different inference stack. Outputs at temperature=0 are typically identical, but token-level scheduling can differ on edge cases (tool-use formatting, very long contexts).
TPM vs RPM – which usually fires first?
Depends on your prompt size. Heavy RAG prompts (8k+ input tokens) hit TPM first; chatty agentic loops hit RPM. Read X-RateLimit-Remaining-Tokens vs Remaining-Requests after the first call to calibrate.
Tool calls work?
Yes – tools=[{type:'function', function:{name,description,parameters}}] then read response.choices[0].message.tool_calls. llama-3.3-70b-versatile is the most reliable; smaller models hallucinate tool names.

Related connectors

See it working on your own data

Everything documented here ships with the platform – try the document tools free, or go live in 7 days.