llm-sdk-provider-gemini 0.2.0

Gemini provider plugin for llm-sdk.

llm-sdk-provider-gemini

Gemini provider plugin for llm-sdk, implemented using google-genai (Vertex AI). This package installs a provider plugin via Python entry points, enabling Gemini support without modifying the core SDK.

---

Table of Contents

• Overview
• Architecture
• Features
• Installation
• Usage
• Configuration
• Project Structure
• Development
• Testing
• Deployment
• Roadmap
• Contributing
• License
• Contact

---

Overview

llm-sdk-provider-gemini is a plugin package that adds support for:

• Gemini chat completions
• Gemini embeddings
• Gemini streaming chat

It integrates with the llm-sdk core using Python entry points, so the provider is automatically discoverable via:

sdk.registry.load_plugins()

---

Architecture

Plugin mechanism

This package declares an entry point:

[project.entry-points."llm_sdk.providers"]
gemini = "llm_sdk_provider_gemini.plugin:GeminiProviderFactory"

At runtime:

1. llm-sdk scans installed packages for llm_sdk.providers
2. Loads the GeminiProviderFactory
3. Calls:
   • spec() for metadata
   • create(settings) to build a provider client

---

Features

• Async Gemini client
• Supports:
  • chat completions
  • embeddings
  • streaming chat
• Uses google-genai (Vertex AI mode)
• Unified SDK error mapping (ProviderError)
• Strong typing
• Clean separation from SDK core

---

Installation

Install core SDK

pip install llm-sdk

Install Gemini provider

pip install llm-sdk-provider-gemini

---

Usage

Async quickstart

import asyncio
from llm_sdk.async_sdk import AsyncSDK


async def main() -> None:
    sdk = AsyncSDK.default()

    # Load installed plugins (entry points)
    sdk.registry.load_plugins()

    resp = await sdk.chat(
        provider="gemini",
        model="gemini-2.5-flash",
        messages=[("user", "Write a haiku about software architecture.")],
    )

    print(resp.content)


if __name__ == "__main__":
    asyncio.run(main())

Streaming example

import asyncio
from llm_sdk.async_sdk import AsyncSDK


async def main() -> None:
    sdk = AsyncSDK.default()
    sdk.registry.load_plugins()

    async for ev in sdk.stream_chat(
        provider="gemini",
        model="gemini-2.5-flash",
        messages=[("user", "Explain AI in Spanish, in 5 short bullets.")],
    ):
        print(ev.delta, end="", flush=True)

    print("\n[done]")


if __name__ == "__main__":
    asyncio.run(main())

Embeddings example

import asyncio
from llm_sdk.async_sdk import AsyncSDK


async def main() -> None:
    sdk = AsyncSDK.default()
    sdk.registry.load_plugins()

    resp = await sdk.embed(
        provider="gemini",
        model="text-multilingual-embedding-002",
        input=["Hola mundo", "Hello world"],
    )

    print(resp.vectors[0][:10])


if __name__ == "__main__":
    asyncio.run(main())

---

Configuration

This provider uses Google credentials from the environment.

Authentication (Vertex AI)

You must have:

• Google Cloud project enabled for Vertex AI
• Proper credentials available through ADC (Application Default Credentials)

Typical local setup:

gcloud auth application-default login

Provider-specific settings

This package typically reads:

• Gemini location (example: us-central1)

Depending on your llm-sdk settings design, you may configure:

export LLM_SDK_LOCATION="us-central1"

Or if you implement namespaced settings:

export LLM_SDK_GEMINI_LOCATION="us-central1"

Project Structure

llm-sdk-provider-gemini/
├─ src/
│  └─ llm_sdk_provider_gemini/
│     ├─ plugin.py
│     ├─ settings.py
│     ├─ async_client.py
│     ├─ sync_client.py
│     └─ __init__.py
├─ tests/
│  ├─ test_factory.py
│  └─ test_contracts.py
├─ pyproject.toml
└─ README.md

---

Setup

Development

python -m venv venv
source venv/bin/activate

pip install -U pip
pip install -e .

Install core SDK in editable mode (local dev)

pip install -e ../llm-sdk

Testing

pytest -q

Notes

Provider tests should include:

• Factory loads correctly
• Spec metadata is correct
• Client implements required methods
• Error mapping works as expected

Avoid real network tests in unit tests. If you want integration tests, put them behind an env flag.

---

Deployment

Build package

python -m build

Upload to TestPyPI

python -m twine upload --repository testpypi dist/*

Upload to PyPI

python -m twine upload dist/*

---

Roadmap

• Better usage extraction for Gemini responses
• Support for Gemini tool calling (if SDK adds tool abstractions)
• Optional support for non-Vertex Gemini API mode
• Optional response caching hooks

---

Contributing

PRs welcome.

Recommended workflow:

1. Fork repo
2. Create a feature branch
3. Add tests
4. Run pytest
5. Submit PR

---

License

MIT License

---

Contact

Author: Esteban Flores