Groq API Inference

Setup

On first use, read setup.md for activation preferences, credential verification, and default workflow setup.

When to Use

User needs to build, integrate, or troubleshoot Groq API inference for chat, tool calling, or speech transcription. Agent handles request shaping, model routing, failure recovery, and safe production patterns.

Architecture

Memory lives in ~/groq-api/. See memory-template.md for structure.

~/groq-api/
├── memory.md           # Status, activation preference, and defaults
├── requests/           # Reusable payload snippets
├── logs/               # Optional debug snapshots
└── experiments/        # Prompt/model A-B notes

Quick Reference

Use these files as decision aids, not as static docs: pick the smallest file that resolves the current blocker.

Core Rules

1. Verify Auth and Endpoint Before Any Work

Check GROQ_API_KEY first and use Authorization: Bearer $GROQ_API_KEY for every request. Use https://api.groq.com/openai/v1 as the base URL and confirm access with /models.

curl -s https://api.groq.com/openai/v1/models \
  -H "Authorization: Bearer $GROQ_API_KEY" | jq '.data[0].id'

2. Start with a Minimal Deterministic Payload

Begin with small prompts and explicit format instructions. Add complexity only after the baseline call is stable.

3. Route by Task, Not by Habit

Use separate model choices for:

• Fast interactive chat
• High-accuracy reasoning
• Speech transcription

Choose from live /models output instead of hardcoding assumptions.

4. Design for Retry and Degradation

For 429 and 5xx, retry with exponential backoff and capped attempts. If a model is overloaded, fail over to a compatible backup model and log the swap.

5. Validate Output Before Downstream Actions

If output feeds code execution or data writes, enforce JSON schema or strict parsing before acting. Reject malformed output early.

6. Treat Speech as a Separate Reliability Path

Speech uploads have different failure modes than chat. Validate input format, check file size, and surface transcription confidence when available.

7. Keep Secrets and User Data Scoped

Never store API keys in files. Keep request logs sanitized and avoid persisting full sensitive prompts unless the user explicitly asks.

Common Traps

• Using stale model IDs copied from old examples -> call /models and select available IDs at runtime.
• Sending giant prompts without truncation -> latency spikes and timeout risk.
• Ignoring 429 backoff guidance -> repeated failures under load.
• Mixing chat and transcription assumptions -> wrong endpoint and payload format.
• Trusting free-form text for automation -> parse and validate before executing.

External Endpoints

All network traffic should be limited to these Groq endpoints for explicit inference tasks requested by the user.

No other data is sent externally.

Security & Privacy

Data that leaves your machine:

• Prompt content sent to Groq inference endpoints
• Audio content sent to Groq transcription endpoint when requested

Data that stays local:

• Workflow preferences in ~/groq-api/memory.md
• Optional local debug notes in ~/groq-api/logs/

This skill does NOT:

• Store GROQ_API_KEY in project files
• Access files outside ~/groq-api/ for persistence
• Call undeclared third-party endpoints
• Modify itself or other skills

Trust

By using this skill, prompts and optional audio content are sent to Groq. Only install if you trust Groq with that data.

Related Skills

Install with clawhub install <slug> if user confirms:

• api — reusable REST patterns, auth, and error handling
• models — model comparison and selection heuristics
• ai — current AI landscape checks before implementation decisions
• fine-tuning — adaptation workflows when prompting is not enough
• langchain — orchestration patterns for multi-step LLM pipelines

Feedback

• If useful: clawhub star groq-api
• Stay updated: clawhub sync