nucleusiq-groq 0.1.0b1

Groq inference provider for the NucleusIQ AI agent framework (official groq Python SDK).

nucleusiq-groq

Groq inference provider for NucleusIQ: Chat Completions via Groq’s OpenAI-compatible API, using Groq’s official groq Python SDK (AsyncGroq / Groq).

Status: 0.1.0b1 (public beta / pre-release). Requires nucleusiq>=0.7.9.

Design, phased roadmap (Phase A vs B), and API caveats: docs/design/GROQ_PROVIDER.md (repo root).

---

Install

PyPI (when published):

pip install nucleusiq nucleusiq-groq

Monorepo (editable):

cd src/providers/inference/groq
uv sync --group dev

Core is pulled via [tool.uv.sources] as an editable path dependency; for pip-only local installs, install nucleusiq from src/nucleusiq first, then this package.

---

Configuration

Variable	Purpose
GROQ_API_KEY	Required. Groq API key.
GROQ_MODEL	Optional. Default for chat/tool examples: llama-3.3-70b-versatile.
GROQ_MODEL_STRUCTURED	Optional. Model for json_schema structured output (example 05 defaults to openai/gpt-oss-20b if unset).

Unsupported Chat Completions fields (e.g. logit_bias, messages[].name) are stripped or rejected at the wire layer; see the design doc.

---

Usage

import asyncio
from nucleusiq.agents import Agent
from nucleusiq.agents.config import AgentConfig, ExecutionMode
from nucleusiq.agents.task import Task
from nucleusiq.prompts.zero_shot import ZeroShotPrompt
from nucleusiq_groq import BaseGroq, GroqLLMParams

async def main() -> None:
    llm = BaseGroq(model_name="llama-3.3-70b-versatile", async_mode=True)
    agent = Agent(
        name="demo",
        prompt=ZeroShotPrompt(),
        llm=llm,
        config=AgentConfig(
            execution_mode=ExecutionMode.DIRECT,
            llm_params=GroqLLMParams(temperature=0.2),
        ),
    )
    await agent.initialize()
    result = await agent.execute(Task(id="1", objective="Capital of France in one short phrase."))
    print(result.output)

asyncio.run(main())

---

Phase A (this beta)

• BaseGroq — call / call_stream, tool calling, structured output (response_format / Pydantic).
• GroqLLMParams — typed, extra="forbid"; merges into provider calls.
• Local function tools — OpenAI-style tool JSON; assistant tool_calls normalized for Groq before each request.
• Retries — rate-limit and transient errors with exponential backoff; 429 + Retry-After on non-stream and streaming open; errors mapped to NucleusIQ LLMError types.
• Hosted tool IDs — nucleusiq_groq.tools.GROQ_COMPOUND_HOSTED_TOOL_IDS / GROQ_GPT_OSS_HOSTED_TOOL_IDS mirror Groq built-in docs for reference only (not wired into call yet).

Groq constraints you should respect: per Structured outputs, streaming and tool use are not currently supported with Structured Outputs on the same request — use non-streaming call for response_format, or skip structured output when streaming / using tools.

Not in Phase A: automatic pass-through for compound_custom, Groq Responses API, and remote MCP (see design doc Phase B).

Integration tests (optional): from src/providers/inference/groq, with GROQ_API_KEY in the environment:

uv run pytest tests/integration -m integration

Default pytest / CI uses -m "not integration" so live calls are optional.

---

Examples

Runnable agents (real API): examples/README.md.

cd src/providers/inference/groq
uv run python examples/agents/01_groq_direct.py

---

Development

From this directory:

uv sync --group dev
uvx ruff check nucleusiq_groq tests examples
uvx pyrefly check
uv run pytest

---

License

MIT — same as the NucleusIQ monorepo.