4 Comments
User's avatar
mucio's avatar

TOON loks like an annotated csv.

It's not clear to me why

```

users[3]{id,name,role}:

1,Alice,admin

2,Bob,user

3,Charlie,admin

```

should be better than

`users.csv`

```

id,name,role

1,Alice,admin

2,Bob,user

3,Charlie,admin

```

But maybe it's LLM magic

Expand full comment
mehdio's avatar

Mutliple things!

As you can see you have a number next to "users". This one can be nested and put next to other blocks.

That means that for each block we have

-name of the object

-number of rows

-field schema

-you know where the nested fields start/end

Expand full comment
Neural Foundry's avatar

The TOON format looks really promisng for reducing token costs, especially for large datasets. I've been strugglign with the same multi-LLM CLI issue and aichat seems like a solid solution. Curious if you've tested the token savings on real-world workloads yet?

Expand full comment
mehdio's avatar

I haven't tested it yet! My personal projects doesn't burn too much cost for me to be worried about that so far but still curious to see if I would see some difference even at lower volume, so will try soon!

Expand full comment