Open Source Summit India 2026

Every token sent to an LLM costs money. When you serialize data as JSON for a prompt, you pay for repeated field names, extra braces, and structural noise on every single row. For large datasets that overhead runs to 40-60% of your token bill & it adds nothing useful to the prompt.

TOON (Token-Oriented Object Notation) is a compact, human-readable format built specifically for LLM prompts. It writes column headers once and streams data as plain rows, similar to CSV, but with full support for nesting, arrays, and schema markers. The result is 40-60% fewer tokens with measurably better LLM accuracy: 73.9% one-shot vs JSON's 69.7% on tabular tasks.
​
This talk covers the TOON format from the ground up: why it exists, how it encodes data, when it wins over JSON and when it does not, and how to use it in real LLM prompts today.

Finally we walk through the toon4s-spark integration, connecting Apache Spark and Databricks to TOON and streaming patterns on Delta Lake.

You will leave knowing exactly how to cut LLM prompt costs, with a format and library you can adopt from any JVM stack today.

Check-
https://github.com/com-vitthalmirji/toon4s
https://toonformat.dev/