Installation

Vane Data uses vane-ai as the distribution package and vane as the Python import package.

Requirements

• Python 3.10 or later.
• A platform with a published Vane Data wheel, or a local C++ build environment for source builds.
• Network and credentials for any remote storage, model provider, or Ray cluster used by your pipeline.

The package metadata installs the required runtime dependencies, including cloudpickle and ray.

Install a released package

Use the released package when a wheel is available for your Python version and platform:

python -m pip install vane-ai

Install the general data-science extras when you need optional dataframe, Arrow, filesystem, or ADBC integration dependencies:

python -m pip install "vane-ai[all]"

The all extra is a convenience group. AI provider libraries are loaded lazily and should be installed for the providers you use:

# Local embedding and classification models
python -m pip install sentence-transformers transformers torch


# Hosted model providers
python -m pip install openai numpy
python -m pip install anthropic
python -m pip install google-genai numpy


# vLLM-backed prompting
python -m pip install vllm

Verify the install

python - <<'PY'
import vane


print("Vane:", vane.__version__)
print("DuckDB:", vane.__duckdb_version__)


con = vane.connect()
con.sql("select 42 as answer").show()
PY

Enable Ray-backed execution

Use Ray only when the workload needs distributed scans, distributed writes, distributed UDFs, or cluster resource placement. For predictable behavior, configure the intended runner before creating connections:

import vane


vane.configure(runner="ray")


con = vane.connect()
con.sql("select 42 as answer").show()

The same setting can be applied with an environment variable:

export VANE_RUNNER=ray

Every Ray worker must be able to import the same Python packages, access the same storage systems, and see any model files or provider credentials required by the pipeline.

Build from source

Build from source when you are developing Vane Data, testing unreleased changes, or using a platform without a matching wheel.

Clone the repository with submodules:

git clone --recursive https://github.com/AstroVela/vane.git
cd vane

If the checkout already exists and submodules are missing:

git submodule update --init --recursive

Install common Debian or Ubuntu build tools:

sudo apt-get update
sudo apt-get install -y build-essential cmake ninja-build pkg-config curl zip unzip tar flex bison

Prepare vcpkg for C++ dependencies:

git clone https://github.com/microsoft/vcpkg.git ../vcpkg
../vcpkg/bootstrap-vcpkg.sh

Install Python build tooling and build in editable mode:

python -m pip install -U pip
python -m pip install cmake ninja scikit-build-core "pybind11[global]"


python -m pip install -e . --no-build-isolation -v \
  --config-settings=cmake.define.CMAKE_TOOLCHAIN_FILE="$PWD/../vcpkg/scripts/buildsystems/vcpkg.cmake"

If an existing checkout was previously configured without the vcpkg toolchain, remove the old CMake build directory before rebuilding so CMake does not reuse an incompatible cache.

Build configuration notes

The source build configuration currently includes these DuckDB extensions:

core_functions;json;parquet;icu;jemalloc;httpfs

This means JSON, Parquet, ICU, jemalloc, and HTTP/S3 filesystem support are part of the configured build. Other DuckDB extensions depend on the build and runtime environment. Verify extension availability in the environment where the job runs.

The project includes wheel-build configuration for Linux manylinux 2.28 on x86_64 and aarch64, macOS, and Windows. Actual package availability depends on the release artifacts published for a given version.

Troubleshooting

• ModuleNotFoundError: No module named 'vane': confirm that vane-ai was installed into the Python environment running your script.
• Missing provider libraries: install the provider package directly, such as openai, anthropic, google-genai, vllm, or sentence-transformers.
• Ray workers cannot import a module: install the same dependencies on every worker or provide a Ray runtime environment that includes them.
• Remote file reads fail: verify the DuckDB extension, storage credentials, endpoint settings, and worker-side environment variables.

Next: SQL quickstart or Python quickstart.