docarray

What it is and what it does

DocArray is a Python library for defining, organizing, and working with multimodal data in machine learning workflows. It provides Pydantic-based schema definitions that let you declare document types with typed fields—including tensors with explicit shapes—and then collect them into vectorized or list-based containers for batch processing. The library integrates with NumPy and other tensor frameworks, and is designed to work seamlessly with web frameworks and microservice platforms.

You use DocArray when you need to represent complex, heterogeneous data (images, text, embeddings, metadata) in a structured way that mirrors how machine learning models consume it. It handles both single documents and bulk collections: DocVec stacks tensors for efficient batch operations, while DocList preserves individual tensor structures for streaming or re-ranking. The library also supports nested document composition and can serialize data as JSON over HTTP or Protobuf over gRPC.

Use it for

• Define typed schemas for multimodal training data with tensor shape validation, then batch them for model training.
• Build API endpoints that accept and return structured multimodal documents with automatic validation.
• Organize and transmit image, text, and embedding data together in a single document structure for neural search applications.
• Compose nested document hierarchies (e.g., a document containing both image and text sub-documents) for complex data pipelines.
• Serialize multimodal collections to JSON or Protobuf for inter-service communication in microservice architectures.

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

docarray on PyPI

Before you install

Low install friction with six runtime dependencies (pydantic, numpy, orjson, typing-inspect, types-requests, rich). Active maintenance with last commit on 2026-03-27 and 3125 repository stars.

License in practice

Licensed under Apache 2.0 (permissive), allowing free use, modification, and distribution in both open-source and commercial projects with minimal restrictions.

Quickstart

pip install docarray

from docarray import BaseDoc, DocVec
from docarray.typing import ImageUrl
import numpy as np

class MyDocument(BaseDoc):
    description: str
    image_url: ImageUrl
    tensor: np.ndarray

vec = DocVec[MyDocument]([
    MyDocument(
        description="A cat",
        image_url="https://example.com/cat.jpg",
        tensor=np.zeros((3, 224, 224)),
    )
])

Package facts

Evidence: docarray-0.41.0-py3-none-any.whl

Tags

More Software Development packages

See also tensordict · tensordict-nightly · tensorly · autoray · keras · safetensors · numpydantic · xarray · torchtyping · keras-nightly