Video Intelligence Pipeline in 20 MinutesBuild a complete video pipeline in one sitting

Make Building MultimodalAI Data Apps Dead Simple

Declare once, run locally, deploy to cloud.

Talk to Us

From the creators of Apache Parquet and Impala and engineers who worked at:

The product loop

Declare → Local → Cloud

Same app logic everywhere — local for speed, cloud for the target.

1Declare

Declare the app in Python

Schema, computed columns, indexes, and HTTP routes in one declarative Python file. Same logic for local and cloud.

2Local

Run locally for fast iteration

pip install and iterate on your laptop. Tables, pipelines, and endpoints without stitching five services.

3Cloud

Deploy the same app to cloud

Bind the same app to a hosted database and Deploy endpoints. No rewrite between environments.

What you declare

One Table to Rule Them All.

Seven primitives unified in one declarative Python table. One import to find them.

Storage

pxt.create_table()

Multimodal Tables

Store images, video, audio, and documents as first-class citizens. No opaque blobs, no external blob storage required.

100% local-firstBuilt-in versioning and time travelLineage to every cell

Orchestration

add_computed_column()

Computed Columns

Declarative IVM. Change a model or UDF, only the affected cells recompute. No DAG file, no orchestrator.

ReplacesAirflow · Prefect · Celery

Retrieval

add_embedding_index()

Embedding Indexes

Vector search built in. Indexes self-maintain as data and models change.

ReplacesPinecone · Weaviate · Qdrant

Storage

create_view()

Views & Iterators

Chunk documents, extract video frames, split audio, slice strings. All declarative.

Replacescustom ETL + chunking scripts

Open source

$pip installpixeltable

Orchestration

@pxt.udf · @pxt.query · pixeltable.functions.*

UDFs, Transforms, and 25+ Providers

Custom functions, built-in transforms, and first-party AI providers in one layer. Batching, retries, rate limits, and caching handled for you.

OpenAIAnthropicGeminiBedrockMistralGroqDeepSeekHugging FaceCLIPWhisperVoyageJinaTogetherFireworksOllamaReplicatefal.aiRunwayMLBFL FLUXTwelve Labs+ more

Serving

pxt.tools() · invoke_tools()

Tool-Calling Agents

Tables as agents. UDFs and @pxt.query as tools. Persistent memory and reasoning traces.

ReplacesLangChain · custom agent glue

Serving

FastAPIRouter · pxt serve

FastAPI Serving

Tables and @pxt.query become HTTP endpoints. Background jobs included.

Replaceshand-written FastAPI + Pydantic

Full API reference

Declarative and incremental

Declare It Once. It Stays Correct.

Computed columns recompute only what changed — your feedback loop never breaks.

Your data is scattered.

Ingest

You need to know where your data is, where it came from, and where it goes, as you build and as you grow. That requires integrity, not more glue.

What you get

The only system where video, audio, images, and documents are first-class column types, not opaque blobs. Schema, versioning, and lineage from the moment data enters. Your files stay where they are.

Replaces

boto3 upload scriptsMetadata DBSync glue code

Learn more

Tables & Data Type System Cloud Storage

1media = pxt.create_table('app.media', {
2  'video': pxt.Video,
3  'doc': pxt.Document,
4  'meta': pxt.Json
5})
6
7media.insert([
8  {'video': 's3://bucket/demo.mp4'},
9  {'doc': '/local/report.pdf'}
10])

Insert a row → every downstream step triggers automatically. One system of record for all modalities.

Your feedback loop breaks.

Process

Every change means re-running pipelines by hand, re-checking drift, and hoping nothing broke. The iteration speed that should be your advantage becomes your bottleneck.

Your infrastructure doesn't compound.

Ship

Shipping is easy. Evolving is hard. Every iteration should make your system smarter. That's the competitive advantage glue code can never provide.

Get Started Contact Us GitHub

Free & open source · Apache 2.0 · No account required

Your Backend for Multimodal AI Data Apps

One system to store, transform, index, serve, and version — declare locally, deploy to cloud.

Instead of stitching together	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`native types with caching
Airflow / Prefect / Celery	`Computed columns`trigger on insert, no orchestrator needed
LangChain / LlamaIndex (RAG)	`@pxt.query + .similarity()`computed column chaining
Hand-written FastAPI + Pydantic	`FastAPIRouter / pxt serve`queries become endpoints in one line
pandas / polars (multimodal)	`.sample(), add_computed_column()`prototype to production
DVC / MLflow / W&B	`history(), revert(), time travel`built-in snapshots
Custom retry / rate-limit / caching	`Built into every AI integration`results cached, only new rows recomputed

One pip install, one Python API, not a stack of services to wire together.

See It In Action

Declare schema and AI workflow as computed columns — run locally or deploy to cloud.

Colab GitHub Starter Kit

Quick Start ▶ Demo

Transactional, incremental, fully versioned: multimodal AI workloads that run themselves

1# Video intelligence: ingest, extract, enrich, index, query
2import pixeltable as pxt
3from pixeltable.functions.video import frame_iterator
4from pixeltable.functions import yolox, gemini, whisper, twelvelabs
5from pixeltable.functions.huggingface import clip, sentence_transformer
6
7# 01 Ingest: native multimodal types
8videos = pxt.create_table('app.videos', {
9    'video': pxt.Video,
10    'title': pxt.String,
11})
12
13# 02 Extract: frames, audio, transcript (automatic on insert)
14frames = pxt.create_view('app.frames', videos,
15    iterator=frame_iterator(video=videos.video, fps=1)
16)
17videos.add_computed_column(audio=videos.video.extract_audio())
18videos.add_computed_column(
19    transcript=whisper.transcribe(videos.audio, model='base')
20)
21
22# 03 Enrich: Gemini multimodal + YOLOX + custom UDF
23@pxt.udf
24def label_scene(detections: list[dict], description: str) -> str:
25    objects = [d['class'] for d in detections[:5]]
26    return f"{description} | objects: {', '.join(objects)}"
27
28videos.add_computed_column(
29    description=gemini.generate_content(
30        [videos.video, 'Describe this video in one sentence.'],
31        model='gemini-2.5-flash'
32    )
33)
34frames.add_computed_column(
35    detections=yolox(frames.frame, model_id='yolox_s')
36)
37frames.add_computed_column(
38    label=label_scene(frames.detections, videos.description)
39)
40
41# 04 Index: CLIP, MiniLM, Twelve Labs, always in sync
42frames.add_embedding_index('frame',
43    image_embed=clip.using(model_id='openai/clip-vit-base-patch32'))
44frames.add_embedding_index('label',
45    string_embed=sentence_transformer.using(model_id='all-MiniLM-L6-v2'))
46videos.add_embedding_index('video',
47    embedding=twelvelabs.embed.using(model_name='marengo3.0'))
48
49# 05 Query: similarity + metadata filtering in one expression
50@pxt.query
51def find_scenes(query_text: str, ref_image: pxt.Image, title: str):
52    text_sim = frames.label.similarity(string=query_text)
53    img_sim = frames.frame.similarity(ref_image)
54    return (frames
55        .where(videos.title.contains(title))
56        .order_by(text_sim + img_sim, asc=False)
57        .limit(10)
58        .select(frames.frame, frames.label)
59    )

app.videostable

Column	Type	Computed With
video	Video
title	String
audio	Audio	extract_audio()
transcript	Json	whisper(base)
description	String	gemini(video)
embedding_index		Twelve Labs (video)

app.framesview → videos

Column	Type	Computed With
frame	Image	frame_iterator
detections	Json	yolox(frame)
label	String	@pxt.udf
embedding_index		CLIP (image)
embedding_index		MiniLM (text)

FastAPIserve

POST/api/videos→ insert

GET/api/search→ find_scenes()

GET/api/frames→ .select().collect()

Insert a video → Gemini, YOLOX, Whisper + custom @pxt.udf as computed columns → CLIP, MiniLM, Twelve Labs indexes → query via @pxt.query. Experiment to production.

Use cases

What Can You Build?

Three paths from day one on the same table abstraction — same pip install.

Data Wrangling for ML

Curate, augment, and export training datasets. Auto-annotate with AI, version everything, export to Parquet or PyTorch.

frames = pxt.create_view('ml.frames', videos,
  iterator=frame_iterator(video=v.video, fps=1))
frames.add_computed_column(detections=yolox(frames.frame))
export_parquet(frames, 'training_data/')

Multimodal IngestAuto-Annotate with AIVersion & SnapshotExport Parquet / Pandas

Learn more

Agents & MCP

Tool calling, persistent memory, and decision traces, not stateless glue. UDFs as LLM tools, MCP servers, semantic recall.

tools = pxt.tools(get_weather, search_docs)
agent.add_computed_column(response=chat_completions(
  messages=msgs, tools=tools))
memory.add_embedding_index('content', string_embed=embed_fn)

@pxt.udf as ToolsMCP Server IntegrationPersistent MemorySemantic Recall

Learn more

End-to-End Multimodal

Video in, finished product out. RAG, semantic search, multimodal APIs: one table, same primitives.

videos.add_computed_column(transcript=whisper.transcribe(audio))
videos.add_computed_column(hook_video=gemini.generate_videos(...))
videos.add_computed_column(final=with_audio(hook_video, tts_audio),
  destination='s3://bucket/output/')