Built-in LLM and Embedding UDFs

Geneva ships pre-built UDFs for common LLM providers so you don’t have to write custom classes for everyday embedding and generation tasks.

Provider	Embeddings	Generation	Runs locally	Install extra
OpenAI	✓	✓	—	geneva[udf-text-openai]
Gemini	✓	✓	—	geneva[udf-text-gemini]
Sentence Transformers	✓	—	✓	geneva[udf-text-sentence-transformers]

OpenAI and Gemini UDFs make remote API calls that incur per-token costs. Sentence Transformers run locally on your workers with no API costs — see GPU acceleration for performance tips.

Comparing models and prompts

Because add_columns accepts a dictionary, you can evaluate multiple models, parameter settings, or prompts in a single pass over your data. Each entry produces its own column, so results sit side by side in the same table for easy comparison.

from geneva.udfs import openai_udf, gemini_udf, openai_embedding_udf

table.add_columns({
    # Compare two embedding models
    "emb_small": openai_embedding_udf(column="body", model="text-embedding-3-small"),
    "emb_large": openai_embedding_udf(column="body", model="text-embedding-3-large"),

    # Compare the same task across providers
    "summary_openai": openai_udf(
        column="body",
        prompt="Summarize in one sentence",
        model="gpt-5-mini",
    ),
    "summary_gemini": gemini_udf(
        column="body",
        prompt="Summarize in one sentence",
        model="gemini-2.5-flash",
    ),
})

This works for any combination — different models from the same provider, different providers, different prompts with the same model, or different dimensionality settings. All columns are computed in parallel during the same backfill job.

To recompute columns later (e.g., after altering a UDF or adding new rows), use backfill:

# Backfill a single column
table.backfill("emb_small")

# Backfill only rows missing a value
table.backfill("summary_openai", where="summary_openai is null")

What’s included

All built-in UDFs share these capabilities:

• API key handling — Keys are captured from your local environment at UDF creation time and securely serialized with the UDF. No cluster-level environment configuration required.
• Retry with backoff — Transient API errors (rate limits, timeouts, server errors) are automatically retried with exponential backoff.
• Batch processing — Embedding UDFs batch multiple rows per API call for better throughput.
• L2 normalization — Embedding UDFs support optional L2 normalization via the normalize parameter (disabled by default since both providers return pre-normalized vectors).

See also

• Working with UDFs — Write custom scalar, batched, and stateful UDFs
• Error handling — Fine-grained retry and skip policies
• Working with blobs — Process binary data (images, audio, video)

API Reference

• Embeddings — sentence_transformer_udf(), gemini_embedding_udf(), openai_embedding_udf()
• Gemini — gemini_udf()
• OpenAI — openai_udf(), openai_embedding_udf()