pixeltable 0.5.23

AI Data Infrastructure: Declarative, Multimodal, and Incremental

The only open source Python library providing declarative data infrastructure for building multimodal AI applications, enabling incremental storage, transformation, indexing, retrieval, and orchestration of data.

Quick Start | Documentation | API Reference | Starter Kit

---

Installation

pip install pixeltable

Demo

https://github.com/user-attachments/assets/b50fd6df-5169-4881-9dbe-1b6e5d06cede

Quick Start

With Pixeltable, you define your entire data processing and AI workflow declaratively using computed columns on tables. Focus on your application logic, not the data plumbing.

# Installation
pip install -qU torch transformers openai pixeltable

# Basic setup
import pixeltable as pxt

# Table with multimodal column types (Image, Video, Audio, Document)
t = pxt.create_table('images', {'input_image': pxt.Image})

# Computed columns: define transformation logic once, runs on all data
from pixeltable.functions import huggingface

# Object detection with automatic model management
t.add_computed_column(
    detections=huggingface.detr_for_object_detection(
        t.input_image,
        model_id='facebook/detr-resnet-50'
    )
)

# Extract specific fields from detection results
t.add_computed_column(detections_text=t.detections.label_text)

# OpenAI Vision API integration with built-in rate limiting and async management
from pixeltable.functions import openai

t.add_computed_column(
    vision=openai.vision(
        prompt="Describe what's in this image.",
        image=t.input_image,
        model='gpt-4o-mini'
    )
)

# Insert data directly from an external URL
# Automatically triggers computation of all computed columns
t.insert(input_image='https://raw.github.com/pixeltable/pixeltable/release/docs/resources/images/000000000025.jpg')

# Query - All data, metadata, and computed results are persistently stored
# Structured and unstructured data are returned side-by-side
results = t.select(
    t.input_image,
    t.detections_text,
    t.vision
).collect()

What Pixeltable Does

When you run the code above, Pixeltable handles storage, orchestration, indexing, versioning, and inference — automatically. Here's the lifecycle:

You Write	Pixeltable Does
pxt.Image, pxt.Video, pxt.Document columns	Stores media, handles formats, caches from URLs
add_computed_column(fn(...))	Runs incrementally, caches results, retries failures
add_embedding_index(column)	Manages vector storage, keeps index in sync
@pxt.udf / @pxt.query	Creates reusable functions with dependency tracking
table.insert(...)	Triggers all dependent computations automatically
t.sample(5).select(t.text, summary=udf(t.text))	Experiment on a sample — nothing stored, calls parallelized and cached
table.select(...).collect()	Returns structured + unstructured data together
(nothing — it's automatic)	Versions all data and schema changes for time-travel

That single workflow replaces most of the typical AI stack:

Instead of …	Pixeltable gives you …
PostgreSQL / MySQL	pxt.create_table() — schema is Python, versioned automatically
Pinecone / Weaviate / Qdrant	add_embedding_index() — one line, stays in sync
S3 / boto3 / blob storage	pxt.Image / Video / Audio / Document types with caching; destination='s3://…' for cloud routing
Airflow / Prefect / Celery	Computed columns trigger on insert — no orchestrator needed
LangChain / LlamaIndex (RAG)	@pxt.query + .similarity() + computed column chaining
pandas / polars (multimodal)	.sample(), ephemeral UDFs, then add_computed_column() to commit — prototype to production
DVC / MLflow / W&B	Built-in history(), revert(), time travel (table:N), snapshots
Custom retry / rate-limit / caching	Built into every AI integration; results cached, only new rows recomputed
Custom ETL / glue code	Declarative schema — Pixeltable handles execution, caching, incremental updates

On top of these, Pixeltable ships with built-in functions for media processing (FFmpeg, Pillow, spaCy), embeddings (sentence-transformers, CLIP), and 30+ AI providers (OpenAI, Anthropic, Gemini, Ollama, and more). For anything domain-specific, wrap your own logic with @pxt.udf. You still write the application layer (FastAPI, React, Docker).

Deployment options: Pixeltable can serve as your full backend (managing media locally or syncing with S3/GCS/Azure, plus built-in vector search and orchestration) or as an orchestration layer alongside your existing infrastructure.

Where Did My Data Go?

Pixeltable workloads generate various outputs, including both structured outputs (such as bounding boxes for detected objects) and unstructured outputs (such as generated images or video). By default, everything resides in your Pixeltable user directory at ~/.pixeltable. Structured data is stored in a Postgres instance in ~/.pixeltable. Generated media (images, video, audio, documents) are stored outside the Postgres database, in separate flat files in ~/.pixeltable/media. Those media files are referenced by URL in the database, and Pixeltable provides the "glue" for a unified table interface over both structured and unstructured data.

In general, the user is not expected to interact directly with the data in ~/.pixeltable; the data store is fully managed by Pixeltable and is intended to be accessed through the Pixeltable Python SDK.

See Working with External Files for details on loading data from URLs, S3, and local paths.

Key Principles

Store: Unified Multimodal Interface

pxt.Image, pxt.Video, pxt.Audio, pxt.Document, pxt.Json – manage diverse data consistently.

t = pxt.create_table(
   'media',
   {
       'img': pxt.Image,
       'video': pxt.Video,
       'audio': pxt.Audio,
       'document': pxt.Document,
       'metadata': pxt.Json
   }
)

→ Type System · Tables & Data

Orchestrate: Declarative Computed Columns

Define processing steps once; they run automatically on new/updated data. Supports API calls (OpenAI, Anthropic, Gemini), local inference (Hugging Face, YOLOX, Whisper), vision models, and any Python logic.

# LLM API call
t.add_computed_column(
   summary=openai.chat_completions(
       messages=[{"role": "user", "content": t.text}], model='gpt-4o-mini'
   )
)

# Local model inference
t.add_computed_column(
   classification=huggingface.vit_for_image_classification(t.image)
)

# Vision analysis
t.add_computed_column(
   description=openai.vision(prompt="Describe this image", image=t.image)
)

→ Computed Columns · AI Integrations · Sample App: Prompt Studio

Iterate: Explode & Process Media

Create views with iterators to explode one row into many (video→frames, doc→chunks, audio→segments).

from pixeltable.functions.video import frame_iterator
from pixeltable.functions.document import document_splitter

# Document chunking with overlap & metadata
chunks = pxt.create_view('chunks', docs,
   iterator=document_splitter(
       document=docs.doc,
       separators='sentence,token_limit',
       overlap=50, limit=500
   ))

# Video frame extraction
frames = pxt.create_view('frames', videos,
   iterator=frame_iterator(video=videos.video, fps=0.5))

→ Views · Iterators · RAG Pipeline

Index: Built-in Vector Search

Add embedding indexes and perform similarity searches directly on tables/views.

t.add_embedding_index(
   'img',
   embedding=clip.using(model_id='openai/clip-vit-base-patch32')
)

sim = t.img.similarity(string="cat playing with yarn")
results = t.order_by(sim, asc=False).limit(10).collect()

→ Embedding Indexes · Semantic Search · Image Search App

Extend: Bring Your Own Code

Extend Pixeltable with UDFs, reusable queries, batch processing, and custom aggregators.

@pxt.udf
def format_prompt(context: list, question: str) -> str:
   return f"Context: {context}\nQuestion: {question}"

@pxt.query
def search_by_topic(topic: str):
   return t.where(t.category == topic).select(t.title, t.summary)

→ UDFs Guide · Custom Aggregates

Agents & Tools: Tool Calling & MCP Integration

Register @pxt.udf, @pxt.query functions, or MCP servers as callable tools. LLMs decide which tool to invoke; Pixeltable executes and stores results.

# Load tools from MCP server, UDFs, and query functions
mcp_tools = pxt.mcp_udfs('http://localhost:8000/mcp')
tools = pxt.tools(get_weather_udf, search_context_query, *mcp_tools)

# LLM decides which tool to call; Pixeltable executes it
t.add_computed_column(
   tool_output=invoke_tools(tools, t.llm_tool_choice)
)

→ Tool Calling Cookbook · Agents & MCP · Pixelbot · Pixelagent

Query & Experiment: The Best Path from Prototype to Production

Unlike pandas/polars, Pixeltable persists everything, parallelizes API calls automatically, caches results, and turns your experiment into production with one line change. No separate notebook → pipeline handoff:

# Explore with a familiar DSL — filter, sample, apply UDFs ephemerally
results = (
   t.where(t.score > 0.8)
   .order_by(t.timestamp)
   .select(t.image, score=t.score)
   .limit(10)
   .collect()
)

# Sample 5 rows and test a UDF — nothing stored, API calls parallelized and cached
t.sample(5).select(t.text, summary=summarize(t.text)).collect()

# Happy? One line to commit — runs on full dataset, skips already-cached rows
t.add_computed_column(summary=summarize(t.text))

→ Queries & Expressions · Iterative Workflow · Version Control

Version: Data Persistence & Time Travel

All data is automatically stored and versioned. Query any prior version.

t = pxt.get_table('my_table')  # Get a handle to an existing table
t.revert()  # Undo the last modification

t.history()  # Display all prior versions
old_version = pxt.get_table('my_table:472')  # Query a specific version

→ Version Control · Data Sharing

Import/Export: I/O & Integration

Import from any source and export to ML formats.

# Import from files, URLs, S3, Hugging Face
t.insert(pxt.io.import_csv('data.csv'))
t.insert(pxt.io.import_huggingface_dataset(dataset))

# Export to analytics/ML formats
pxt.io.export_parquet(table, 'data.parquet')
pytorch_ds = table.to_pytorch_dataset('pt')  # → PyTorch DataLoader ready
coco_path = table.to_coco_dataset()          # → COCO annotations

# ML tool integrations
pxt.create_label_studio_project(table, label_config)  # Annotation
pxt.export_images_as_fo_dataset(table, table.image)   # FiftyOne

→ Data Import · PyTorch Export · Label Studio · Data Wrangling for ML

Tutorials & Cookbooks

Fundamentals	Cookbooks	Providers	Sample Apps
All →	All →	22 providers →	All →

External Storage and Pixeltable Cloud

Supported storage providers:

Store computed media using the destination parameter on columns, or set defaults globally via PIXELTABLE_OUTPUT_MEDIA_DEST and PIXELTABLE_INPUT_MEDIA_DEST. See Configuration.

Data Sharing: Publish datasets to Pixeltable Cloud for team collaboration or public sharing. Replicate public datasets instantly—no account needed for replication.

import pixeltable as pxt

# Replicate a public dataset (no account required)
coco = pxt.replicate(
    remote_uri='pxt://pixeltable:fiftyone/coco_mini_2017',
    local_path='coco-copy'
)

# Publish your own dataset (requires free account)
pxt.publish(source='my-table', destination_uri='pxt://myorg/my-dataset')

# Store computed media in external cloud storage
t.add_computed_column(
    thumbnail=t.image.resize((256, 256)),
    destination='s3://my-bucket/thumbnails/'
)

Data Sharing Guide | Cloud Storage | Public Datasets

Built with Pixeltable

Project	Description
Starter Kit	Production-ready FastAPI + React app with deployment configs for Docker, Helm, Terraform (EKS/GKE/AKS), and AWS CDK
Pixelbot	Multimodal AI agent, an interactive data studio with on-demand ML inference, media generation, and a database explore
Pixelagent	Lightweight agent framework with built-in memory and tool orchestration
Pixelmemory	Persistent memory layer for AI applications
MCP Server	Model Context Protocol server for Claude, Cursor, and other AI IDEs

Contributing

We love contributions! Whether it's reporting bugs, suggesting features, improving documentation, or submitting code changes, please check out our Contributing Guide and join the Discussions or our Discord Server.

License

Pixeltable is licensed under the Apache 2.0 License.