What is Pixeltable?

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

# LLM API call
t.add_computed_column(summary=openai.chat_completions(
    messages=[{'role': 'user', 'content': 'Summarize: ' + t.text}]
))

# Local model inference
t.add_computed_column(objects=yolox(t.image, model_id='yolox_s'))

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

# Extract frames from video at 1 fps
frames = pxt.create_view('myapp.frames', t, iterator=frame_iterator(t.video, fps=1))

# Chunk documents for RAG
chunks = pxt.create_view('myapp.chunks', t, iterator=document_splitter(t.document))

t.add_embedding_index('text', embedding=openai.embeddings())

# Search by similarity
results = t.order_by(t.text.similarity('find relevant docs'), asc=False).limit(10)

@pxt.udf
def extract_entities(text: str) -> list[str]:
    # Your custom logic
    return entities

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

# Load tools from MCP server, UDFs, and queries
mcp_tools = pxt.mcp_udfs('http://localhost:8000/mcp')
tools = pxt.tools(search_by_topic, extract_entities, *mcp_tools)

# LLM decides which tool to call; Pixeltable executes it
t.add_computed_column(response=openai.chat_completions(
    messages=[{'role': 'user', 'content': t.question}],
    tools=tools
))
t.add_computed_column(result=openai.invoke_tools(tools, t.response))

# Query data with familiar syntax
results = t.where(t.score > 0.8).order_by(t.timestamp).limit(10).collect()

# Test transformations on sample rows BEFORE adding to table
t.select(t.text, summary=summarize(t.text)).head(3)  # Nothing stored yet
t.add_computed_column(summary=summarize(t.text))      # Now commit to all rows

t.history()                    # View all versions
t.revert(version=5)            # Rollback changes
old_data = pxt.get_table('myapp.media:3')  # Query past version

# 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 ML/analytics formats
pxt.io.export_parquet(t, 'output.parquet')
loader = DataLoader(t.to_pytorch_dataset(), batch_size=32)
coco_path = t.to_coco_dataset()

pxt.publish(t, 'my-dataset')              # Share publicly
pxt.replicate('user/dataset', 'local')   # Pull to local