langchain-apertis

Native LangChain integration for the Apertis AI API. It uses the official apertis SDK directly and does not depend on langchain-openai.

Installation

pip install -U langchain-apertis
export APERTIS_API_KEY="..."

Chat completions

from langchain_apertis import ChatApertis

llm = ChatApertis(model="gpt-5.5", temperature=0)
print(llm.invoke("Explain what a native LangChain provider is.").text)

bind_tools() and with_structured_output() use Apertis' OpenAI-compatible tool format. A current Web Search model keeps its exact :web model suffix and exposes source records safely in response_metadata["apertis"]["web_sources"].

search = ChatApertis(
    model="gpt-5.5:web",
    web_results_count=5,
    web_content_length="medium",
)
response = search.invoke("What changed in this week's AI news?")
print(response.response_metadata["apertis"]["web_sources"])

Both current SDK Web Search stream-status shapes are exposed before model text. When the SDK supplies a status string, stream() and astream() retain it and set response_metadata["apertis"]["web_search_status"] to "in_progress".

Multimodal and audio chat

ChatApertis forwards LangChain's OpenAI-shaped image, input-audio, and video content parts directly to the official SDK. Model availability remains account- and model-dependent; the integration preserves the request contract rather than claiming a static catalog.

from langchain_core.messages import HumanMessage

vision = ChatApertis(model="gpt-5.5")
response = vision.invoke(
    [
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image."},
                {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
            ]
        )
    ]
)

For audio-output requests, use the Apertis SDK-compatible modalities and audio parameters. Returned audio data is retained in response.additional_kwargs["audio"]; Web Search URL citations are retained in response.response_metadata["apertis"]["annotations"].

The current SDK's response_format and extra_body request fields are also direct ChatApertis constructor parameters.

Development verification

The public package metadata resolves published apertis>=0.3.0 from PyPI.

uv sync --group dev
uv run pytest -m "not integration"
uv build
docker build -t langchain-apertis-test -f Dockerfile.test .
docker run --rm --network none langchain-apertis-test

Publishing

After apertis>=0.3.0 is available on PyPI, configure PyPI Trusted Publishing for apertis-ai/langchain-apertis, .github/workflows/publish.yml, and the pypi environment. The manual Publish to PyPI workflow uploads only when dispatched with the publish confirmation value.