Groq integration

What it does

Groq serves open models on custom LPU hardware built for low-latency inference. Connect a Groq account and your workflows can generate chat completions and structured JSON, run the stateful Responses API, transcribe or translate audio, synthesise speech, and manage the files and batch jobs behind large asynchronous workloads. Every model field is a live picker backed by the connected account's own roster, so your clients only ever see models they can actually call.

Connect a Groq account

Groq authenticates with a single API key.

1. Sign in at console.groq.com. The free tier needs no card.
2. Create a key at console.groq.com/keys and copy it. Groq shows the value once, but you can mint a replacement at any time.
3. In TaskJuice, add a Groq node to a workflow, open its connection picker, and paste the key.

Keys are unscoped, so a single key reaches every action your account's plan allows.

Triggers

Groq publishes no webhook or event surface, so this integration ships actions only. To react to a Groq result, put a Groq action downstream of whichever trigger starts the work.

Actions

Chat and reasoning

• groq/create-chat-completion — Generate a chat completion, with tool calling, structured outputs and JSON mode.
• groq/create-response — Generate a response through the Responses API, including multi-turn conversations continued by ID.

Models

• groq/list-models — List every model the connected account can use, with context window, modalities and pricing.
• groq/get-model — Retrieve one model's metadata.

Audio

• groq/create-transcription — Transcribe speech in an audio or video file, optionally with segment and word timestamps.
• groq/create-translation — Translate speech in an audio or video file into English text.
• groq/create-speech — Convert text into spoken WAV audio.

Files

• groq/upload-file — Upload a JSONL file for batch processing or fine-tuning.
• groq/list-files — List uploaded files, optionally filtered by purpose.
• groq/get-file — Retrieve one file's metadata.
• groq/download-file-content — Download a file's contents, including batch output.
• groq/delete-file — Permanently delete an uploaded file.

Batches

• groq/create-batch — Start an asynchronous batch job from an uploaded JSONL file at a 50 percent discount.
• groq/list-batches — List batch jobs on the account.
• groq/get-batch — Check a batch job's status and collect its output file ID.
• groq/cancel-batch — Cancel an in-flight batch job.

Fine-tuning (deprecated)

Groq offers LoRA fine-tuning only to Enterprise-tier customers under a sales agreement, so the four fine-tuning actions (create-fine-tuning, list-fine-tunings, get-fine-tuning, delete-fine-tuning) are deprecated. Workflows that already use them keep running; they can no longer be added to new workflows.

Known limitations

Files and batches need the Developer plan. On the free tier those nine actions answer 403 not_available_for_plan. The Developer plan is pay-per-token with no monthly fee; once the connected account is on it the actions work with no change in TaskJuice.

Text-to-speech models need a one-time terms acceptance. The first Create Speech call on a new organisation returns 400 model_terms_required. An org admin accepts the model's terms once in the Groq playground, after which the action runs normally. Each model has its own voices — Orpheus V1 English offers autumn, diana, hannah, austin, daniel and troy — and an unknown voice returns 400 listing the valid ones.

Only whisper-large-v3 can translate. Create Transcription works with either Whisper model, but Create Translation is rejected by whisper-large-v3-turbo with "does not support translate". Groq's models API exposes no capability flag that distinguishes them, so the picker lists both and the choice is yours.

Transcription timestamps are one granularity per call. Choose either segment or word, and set Response Format to verbose_json to receive them.

Model IDs change. Groq retires models as it adds them, and a retired ID returns 404 model_not_found. Model fields are pickers bound to the live roster for exactly this reason, so prefer selecting a model over typing or expression-binding a fixed ID.

Free-tier rate limits are low. Expect roughly 30 requests per minute and 14,400 requests per day at the organisation level, varying by model. Groq signals throttling with 429, which TaskJuice retries. Adding a card lifts the limits substantially.

Speech input is capped. Transcription and translation accept audio up to Groq's per-request file-size ceiling. For longer recordings, split the audio upstream or move the work to a batch job.