Make Building Multimodal AI Data Apps Dead Simple

The unified multimodal backend for AI data apps in one Python file.

Open in your agent

Cursor Claude ChatGPT

Reading this as an agent? Start at /get-started.md

Declare → Experiment → Serve

Pixeltable is the database, orchestration, and serving layers.

pxt schema update

Declare

Creates the tables app.py declares. Media and structured columns live together.

pxt dashboard

Experiment

Insert a sample. Query. Change a column. Assigned columns run on insert.

pxt service update

Serve

Starts the FastAPIRouter services app.py declares. Local dir or pxt://. pxt service list prints the URL. pxt service run is local only.

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

From the creators of Apache Parquet and Impala.

Insert a row, the rest follows

RAG that stays in sync

class Chunks(TableModel, iterator=document_splitter(Docs.document))

Insert a document. Chunks, embeddings, and answers update. Retrieval is an @pxt.query on that view.

RAG pipeline

Media that processes on insert

class Frames(TableModel, iterator=frame_iterator(video=Videos.video, fps=1))

Insert a video or image. Frames, thumbnails, and captions are the view.

Video frames

Agents you can query later

tool_output = invoke_tools(tools, response)

Insert a message. Tool calls and the answer are assignments, not a loop in memory.

Tool calling

HTTP without writing handlers

api.add_insert_route(Docs, path='/docs', inputs=[Docs.document])

add_insert_route is the insert URL. Keep @router.post for multi-table work or side effects.

Serving

One app.py

No extra blob store, vector database, orchestrator, or glue. The walkthrough is the receipt.

app.py

terminal

Subclass TableModel. Every annotation becomes a stored column.

import pixeltable as pxt

TableModel = pxt.model_base()


class Videos(TableModel, name='videos'):
    video: pxt.Video  # also pxt.Image, pxt.Audio, pxt.Document, pxt.Json
    title: pxt.String

After this step

videos table

videoVideo

titleString

GitHub repo Why AI transformations belong in the schema Deployment patterns

After launch, the same app.py

Only new rows recompute

label = label_scene(detections, Videos.description)

Adding a hundred videos costs a hundred videos of work, not the whole table.

Schema changes ship as a reviewed plan

pxt schema diff app.py my_app

Dropping a column or an index needs --allow-destructive. In CI, a pending change exits 2.

Failed rows stay in the table

pxt errors my_app/frames --col label

The cell is None plus errortype and errormsg. Fix the cause and recompute those rows.

Any version is recoverable

pxt revert my_app/frames --steps 3

pxt history shows what a revert would undo before you run it.

Production operations Version control and lineage Browse media and errors with pxt dashboard

Build and maintain are the same job

The backend is one schema you keep, not a stack you assemble.

10-min quickstart View on GitHub

Video, image, audio, and documents are column types, not paths. See solutions