Skip to main content
vane
StarContact us
Vane Data

Overview

Vane Data is a high-performance, multimodal-native data engine for AI workloads. Built on a fork of DuckDB, it extends the core execution engine with native multimodal processing and a unified framework for local and distributed execution.

Start with the Quickstart to process local images with Vane, without a GPU, model API key, or external cluster.

Multimodal InputsOutputs / Outcomes
Sensors
Tables
Documents
Images
Video
Audio
Events
Embeddings

Vane Data

Native Multimodal Data Model
Compute + Inference Operator Graph
Parallel CPU-GPU-IO Execution
Local & Ray Execution
IngestParseTransformInferEnrichPackage
Training Data
Features & Embeddings
Searchable Records
Analytical Results

Vane Core

Local Runtime+Ray Runtime

Unified Multimodal Data Type

File references, decoded images, and tensors in SQL and Python.

Streaming + Backpressure + Dynamic Batching

Continuous flow for large objects with adaptive batching and pressure control.

Overlapped Heterogeneous Execution

CPU, GPU, IO, and model inference overlap through asynchronous scheduling.

Local & Distributed Execution

Run pipelines locally or on distributed Ray clusters.

Native multimodal types

File references and decoded values are distinct types that you can query and transform with SQL or Python:

TypeMeaning
FILEImmutable file reference with location and metadata
IMAGEFILE, AUDIOFILE, VIDEOFILEReferences to encoded image, audio, and video files
IMAGEDecoded image pixels
TENSORNumeric arrays, including audio waveforms

Use built-in media functions to inspect metadata, decode and resize images, resample audio, or extract video frames. Media functions use Python backends by default, with the corresponding vane-ai[image], vane-ai[audio], or vane-ai[video] dependencies. Optional native acceleration uses the experimental native_media extension and its separately installed provider wheel.

Key features

  • Python and SQL interfaces — Build pipelines with DuckDB SQL or the Python Relation API, and extend them with Python UDFs.
  • Built-in AI operations — Invoke models and generate embeddings with your chosen provider, alongside data transformations.
  • Heterogeneous execution — Overlap CPU, GPU, I/O, and model inference workloads through asynchronous scheduling.
  • Local and distributed execution — Run pipelines locally or on Ray clusters, with dynamic batching and backpressure control for varying data sizes and processing costs.

Connect your data

IntegrationSupported operations
Lake table formats: Iceberg, Paimon, DuckLakeDistributed queries and writes
Columnar formats: Lance, VortexDistributed queries and writes; Lance also supports vector and hybrid search
Vector databases: Milvus, QdrantDistributed upserts of query results
Analytical database: DorisDistributed Arrow batch writes

Lake and columnar format extensions are installed as provider packages released independently of vane-ai. Database sinks use their corresponding optional dependencies; follow the linked guides for setup and supported operations.

Next steps

Edit this page