pixeltable-workflows

name	pixeltable-workflows
description	Build Pixeltable workflows for multimodal data processing — declarative tables with computed columns, embeddings, and LLM transformations.
version	0.1.0
author	Jero (LATTICE / MARPA Design Studios)
triggers	["pixeltable","create a pixeltable","multimodal data","computed columns","embedding pipeline"]
tools	["Bash","Read","Write","Edit","Glob","Grep"]

Pixeltable Workflows

USE WHEN the user wants to create Pixeltable tables, build multimodal data pipelines, add computed columns with LLM transformations, or manage embeddings and vector search.

What It Does

Build declarative data workflows using Pixeltable's computed column model:

1. Create Tables — Define tables with typed columns for text, images, video, audio, documents
2. Computed Columns — Add columns that auto-compute via LLM calls, embeddings, or transformations
3. Embeddings — Generate and index vector embeddings for semantic search
4. Queries — Run similarity search, filtering, and aggregation over multimodal data
5. Versioning — Pixeltable tracks all data versions automatically

Usage

create a pixeltable for storing code snippets with embeddings
add an LLM summary column to the documents table
search for similar code patterns using embeddings

Table Creation

import pixeltable as pxt

# Create a directory (namespace)
pxt.create_dir("graftkit")

# Create a table
t = pxt.create_table("graftkit.rootstocks", {
    "name": pxt.String,
    "plan_content": pxt.String,
    "created_at": pxt.Timestamp,
})

# Add a computed column (LLM-powered)
t.add_computed_column(
    summary=pxt.functions.openai.chat_completions(
        messages=[{"role": "user", "content": t.plan_content}],
        model="gpt-4o-mini",
    )
)

# Add embedding column for search
t.add_embedding_index("plan_content", string_embed=e5_embed)

Common Patterns

Pattern	Description
Code indexing	Store source files with AST-extracted metadata + embeddings
Plan analysis	Parse Rootstocks, extract tasks, compute complexity scores
Agent traces	Index agent outputs for similarity-based retrieval
Doc search	Embed documentation for semantic Q&A

Storage

Pixeltable stores data in .pixeltable/ (gitignored). The directory is regenerable from source data.