I show how I replaced JSON with TOON in LLM prompts and saved 40% of tokens
Hello! My name is Andrey, I am a front-end developer at Cloud.ru . I work with the LLM API every day. And every day I send structured data into the context: lists of products, logs, users, metrics. It…

Hello! My name is Andrey, I am a front-end developer at Cloud.ru. I work with the LLM API every day. And every day I send structured data into the context: lists of products, logs, users, metrics. It's all JSON. And all this is money.
At some point I counted how many tokens from my prompt were spent on curly braces, quotes and repeating keys. It turned out - a lot. Too much.
Then I tried TOON. Here's what came out of it.
Read more