AI Functions

Vane Data AI functions turn common model operations into relation methods. They produce embeddings, labels, prompts, and structured responses as table columns while keeping the same execution controls used by UDFs.

Use AI helpers when the operation matches a built-in provider pattern. Use custom UDFs when you need full control over model loading, batching, retries, output shape, or provider behavior.

Relation methods

The relation methods are:

• rel.embed_text(...)
• rel.classify_text(...)
• rel.prompt(...)

The functional API is also available from vane.ai and vane.ai.functions.

AI helpers return the configured output column. If downstream stages also need source columns such as IDs or text, use a custom UDF that returns the complete schema, or explicitly combine the helper output with source rows in a validated step that preserves row counts and ordering.

Provider support

Provider libraries are loaded lazily. Install the libraries for the providers you use.

Text embeddings

Use embed_text to produce a vector column from a text column.

embedding_only = rel.embed_text(
    "text",
    provider="transformers",
    model="sentence-transformers/all-MiniLM-L6-v2",
    output_column="embedding",
    execution_backend="subprocess_actor",
)

The helper writes embeddings as a FLOAT[] output column.

Optional text chunking parameters are available for long inputs:

• max_chunk_chars
• chunk_overlap_chars

Text classification

Use classify_text for zero-shot text labels.

label_only = rel.classify_text(
    "text",
    labels=["invoice", "claim", "other"],
    provider="transformers",
    output_column="label",
)

The output column is VARCHAR.

Prompting

Use prompt to generate a response column from a text prompt column.

response_only = rel.prompt(
    "prompt",
    provider="openai",
    output_column="response",
)

The output column is VARCHAR.

Multimodal prompts

prompt accepts image_columns for rows that include image data, when the selected provider and model support multimodal input.

answer_only = rel.prompt(
    "question",
    image_columns=["image"],
    provider="anthropic",
    output_column="answer",
)

Provider and model capabilities differ. Validate the chosen provider and model with a small sample before running a large job.

Structured output

prompt accepts return_format for providers that support structured output.

from pydantic import BaseModel


class Decision(BaseModel):
    label: str
    reason: str


decision_only = rel.prompt(
    "prompt",
    provider="openai",
    return_format=Decision,
    output_column="decision_json",
)

The relation output column is a JSON string. Treat it as a table column: validate it, parse it downstream if needed, and keep the raw prompt input available for audit.

Provider dependencies

Install provider libraries directly:

# Local embedding and classification models
python -m pip install sentence-transformers transformers torch


# Hosted model providers
python -m pip install openai numpy
python -m pip install anthropic
python -m pip install google-genai numpy


# vLLM-backed prompting
python -m pip install vllm

Execution controls

AI helpers accept execution_backend, and provider options can influence the UDF settings passed to map_batches.

Common controls include:

• batch_size
• concurrency
• max_api_concurrency for provider API calls where supported
• provider-specific GPU options, such as gpus_per_actor for the vLLM provider
• provider-specific retry and error behavior

Example:

embedding_only = rel.embed_text(
    "text",
    provider="openai",
    execution_backend="subprocess_actor",
    concurrency=8,
)

Use provider API concurrency carefully. Increasing concurrency can improve throughput, but it can also hit rate limits or amplify failures.

Production guidance

• Filter and project input columns before calling AI helpers.
• Keep prompt input, model/provider name, and output column names explicit.
• Start with a small sample and inspect outputs before scaling.
• Use actor backends when model initialization is expensive.
• Use Ray only after the local execution path is correct.
• Store model outputs in typed columns and keep enough metadata for debugging.