Ondine

Batch-process your DataFrames with LLMs, without the boilerplate.

Agents reason row-by-row. Ondine computes columns.

In a measured benchmark processing 100,000 rows (DeepSeek Chat, batch size 15, projected from a 30-row real-API sample), Ondine made ~15x fewer API calls than a naive per-row loop (6,666 vs 100,000) and cost ~35% less ($0.4815 vs $0.7411).

Ondine is an open-source Python SDK for data engineers and ML practitioners who need to process large CSV/DataFrame datasets with LLMs. Multi-row batching, prefix caching, and budget controls are built in from day one, not bolted on later.

Quick Start

from ondine import enrich

df = enrich(
    "products.csv",
    "Classify this product into a category: {name} - {description}",
    output_columns=["category"],
    model="gpt-4o-mini",
)

Why Ondine?

Most teams processing tabular data with LLMs hit the same three walls: API cost spirals, brittle pipelines that crash halfway through, and hallucinated outputs that are hard to catch at scale. Ondine is built to solve exactly those three problems.

Problem	Ondine's answer
API bills too high	Multi-row batching (100 rows/call) + prefix caching -- ~35% lower cost in our benchmark
Pipeline crashes lose progress	Checkpointing -- resume from last saved row, not row 0
LLM makes things up	Context Store -- ground each response against your source data
Too many providers to juggle	100+ providers via LiteLLM, single unified API

Key Features

• enrich() -- one-call front door; input type in, same type out
• plan() -- draft a pipeline from a plain-language goal, inspect and approve before execution
• Builder API -- Full control over pipeline configuration
• MCP server (ondine-mcp) -- expose pipeline runs as MCP tools for agent clients, budget-capped by design
• Provider Batch API mode -- OpenAI/Anthropic native Batch jobs for ~50% provider-advertised savings
• 100+ LLM providers via LiteLLM (OpenAI, Anthropic, Groq, Azure, local MLX)
• Multi-row batching -- Process 100 rows per API call
• Prefix caching -- reduces token cost on repeated prompts
• Cost estimation -- Know the cost before you run
• Budget limits -- Hard caps on spending
• Checkpointing -- Resume failed pipelines from where they stopped
• Structured output -- Pydantic models, JSON, regex parsing
• Observability -- OpenTelemetry, Langfuse, structured logging

Installation

pip install ondine

Links

• GitHub Repository
• PyPI Package
• Contributing Guide