Rémi

We changed the way we handle multimodal inputs in Outlines.

Most libraries will give the prompt a special role, separate it from the other inputs. We don’t anymore.

📝 I have been meaning to write this for a while now, but after reading @remilouf.bsky.social's post on the potential of structured generation for LLMs and a post on SerDes debt deriving from Pydantic usage, I decided I should not wait any longer. So here is my new blog post:

🦀 Do you like Rust?
🇫🇷 Do you like Paris?

Then you'll love the conference, Rust In Paris, that my colleague Yvan Sraka is delivering his talk, Beyond Procedural Macros: A Trait-Based Framework for Bindgen, at next month!

www.rustinparis.com

This was absolutely my favorite talk from today! Love it / scared of it

Outlines 0.1.9 release is out! Small release, but some highlights:

- First contribution by @chicham!
- You can now use enums in outlines.generate.choice
- Compilation time improvements.
- Using the rust-native JSON-schema to regex, rather than the Python implementation

This is actually good to show skeptics of structured generation @dottxtai.bsky.social. Demonstrates very clearly that you can stop the model from fighting the constraints by telling it how it’s constrained, even if those constraints might seem absurd (like only using words from the King James Bible)

HF has done, and keeps doing, a lot of good in the AI ecosystem. I’ll die on this hill.

More insults under an apology post on Bluesky than under any outrageous post on Twitter. Be the change you want to see in the world or something like that.

Check out our new YouTube video. We're trying out kind of an interview format, so let me know if you love/hate this

Looking at the comments of this one, I think this is actually the Bad Place

The comments on this are unnecessarily hateful. I thought this place was better.

Cool! Are you able to use the crate, or did you have to port the code?

Give @dottxtai.bsky.social a follow for quality structured generation content!

Agreed, though @willkurt.bsky.social obviously got closer to the optimal prompt for structured than the one the authors used :)

A new paper, "Let Me Speak Freely" has been spreading rumors that structured generation hurts LLM evaluation performance.

Well, we've taken a look and found serious issue in this paper, and shown, once again, that structured generation *improves* evaluation performance!

That’s possible

The "Let Me Speak Freely" got way too much traction too easily, despite a few serious issues.

Well aware of Brandolini's law we published a thorough rebuttal of the main conclusion, highlighted the key methodological flaws and taught a few things in the process.

Hope you enjoy it!

We don’t have to generate code in JSON. With our method you can make the model generate the code in a code block like they suggest in the paper.

A graph showing that structured generation performs better than unstructured generation.

Our new blog post is out!

@willkurt.bsky.social provides a rebuttal for a reasonably well known paper which concluded that structured generation with LLMs always resulted in worse performance.

We do not find the same thing.

blog.dottxt.co/say-what-you...

YouTube video by dottxt Build an AI receipt scanner with vision language models!

Y'ALL this is my first @dottxtai.bsky.social video to break 1k views!!!!!

youtu.be/_SK2sO0_DEY?...

What’s the best LLM observability tool today?

The kind of echo chamber that I like

👋 New followers

I’m Rémi, co-founder and CEO of @dottxtai.bsky.social. We do structured generation with LLM, which means we are able to enforce constraints on the output (regex, JSON Schema, Grammar) during inference time. It’s very cool.

dottxt.co

I’m also finding again the ML content that made me love Twitter at the beginning.

I wake up every day with 100 new followers now. The Starter Packs were a brilliant idea ✨

Hello Bluesky! We're here now. Wonderful to be in the Good Place.

We'll be posting demos, interesting tidbits, and boosting cool AI stuff on here.

dottxt.co

(Reposting from X)
I did a tiny side project: a program that automatically generates a properly formatted .ics file from a text description (e.g. copy/paste an email) using an LLM, and imports it to your favorite calendar app!
Structured generation is cool -- shoutout @remilouf.bsky.social and .txt

demos/earnings-reports at main · dottxt-ai/demos Contribute to dottxt-ai/demos development by creating an account on GitHub.

Go check out our earnings report cookbook:

dottxt-ai.github.io/outlines/main/…

Or, if you want more detail, check out the full demo:

github.com/dottxt-ai/de...

📊 One weird trick to parse earnings reports!

Wall Street analysts HATE this!

Use LLMs to convert messy SEC filings into clean CSV data using structured generation.

The one weird trick? Convert your HTML to markdown + use Outlines.

Link below 👇

demos/its-a-smol-world at main · dottxt-ai/demos Contribute to dottxt-ai/demos development by creating an account on GitHub.

Code: github.com/dottxt-ai/de...