Migrate to Pixeltable

Hosted catalog: Cloud. This page maps the stack you already run onto Pixeltable. Defining the class does not create the table. Run pxt schema update app.py my_app first (Quickstart); the snippets assume that already ran.

Pinecone and Milvus
Postgres
LangGraph
Scripts and Airflow

You do not need a separate vector database. The embedding index sits on the column. Insert keeps it current.

import pixeltable as pxt
import pixeltable.functions as pxtf

TableModel = pxt.model_base()


class Items(TableModel, name='items'):
    body: pxt.String
    __indexes__ = [
        pxt.EmbeddingIndex(
            body,
            embedding=pxtf.huggingface.sentence_transformer.using(
                model_id='sentence-transformers/all-MiniLM-L6-v2'
            ),
            name='body_idx',
        )
    ]

items = pxt.get_table('my_app.items')
sim = items.body.similarity(string='application file')
items.order_by(sim, asc=False).limit(5).select(items.body)

Embedding indexes

You do not need a separate RDBMS for application rows. Typed columns, insert, and select live in the catalog.

import pixeltable as pxt

TableModel = pxt.model_base()


class Orders(TableModel, name='orders'):
    sku: pxt.String
    qty: pxt.Int
    note: pxt.String | None
    __indexes__ = [pxt.BtreeIndex(sku)]

orders = pxt.get_table('my_app.orders')
orders.insert([{'sku': 'a1', 'qty': 2, 'note': None}])
orders.where(orders.qty > 1).select(orders.sku, orders.qty).collect()

If another system must keep a SQL copy, export with export_sql (example on Self-hosting).

You do not need StateGraph, ToolNode, or a checkpointer. After pxt schema update, insert {'message': '...'} and read response and tool_output.

import pixeltable as pxt
import pixeltable.functions as pxtf

TableModel = pxt.model_base()


@pxt.udf
def get_weather(city: str) -> str:
    return f'Weather in {city}: 72°F, sunny'


tools = pxt.tools(get_weather)


class Assistant(TableModel, name='assistant'):
    message: pxt.String
    response = pxtf.openai.chat_completions(
        messages=[{'role': 'user', 'content': message}],
        model='gpt-4o-mini',
        tools=tools,
    )
    tool_output = pxtf.openai.invoke_tools(tools, response)

Tool calling cookbook

You do not need a frame loop, cron, or a full re-run. Insert a video. The iterator adds a frame column on Frames; thumb is a computed column from that frame.

import pixeltable as pxt
import pixeltable.functions as pxtf

TableModel = pxt.model_base()


class Videos(TableModel, name='videos'):
    video: pxt.Video


class Frames(
    TableModel,
    name='frames',
    base=Videos,
    iterator=pxtf.video.frame_iterator(Videos.video, fps=1),
):
    # iterator output
    thumb = frame.resize((256, 256))  # type: ignore[name-defined]

Extract video frames

After pxt schema update, insert with table.insert([...]) as in the snippets above, or POST after pxt service update (Quickstart).

Quickstart

Write app.py, run pxt schema update, then insert from Python or HTTP.

HTTP serving

Insert, compute, update, delete, query, uploads, background jobs.

Getting started

Tools

How to

Resources

Quickstart

HTTP serving