We built what OpenAI didn't.

Optional fields. Unions. Dynamic keys. Composition. Features that matter when you deploy. 🧵 👉 Optional fields

Real-world responses are often partially known or progressive. Optional properties let you express “include it when available” without workarounds.

Say you're parsing articles. Sometimes you get citations. Sometimes you don't:

You may have heard the term "agent" quite a bit. But what does that actually mean?

A few themes of agents:

- Autonomy (operate without human interaction)
- Perception (receives information through sensors)
- Goal-oriented (specific task to accomplish)

Isn't this a program? 👇 Autonomy:

Software is automation by definition -- that's why we build it. Software makes it easy to automate any task, like tracking documents, doing math, embed business logic, construct websites, etc.

Agents do the same, but with different capacities + flexibility .

Introducing STRESSED (STRuctured Generation Security System Evaluating Data)!

STRESSED is an anxious security intern that monitors your logs (system/web/etc) for security risks, and provides a simple overview of what's going on in your system using structured generation. Go check out the demo code here! Play around with it and let us know what you think.

github.com/dottxt-ai/demo…

Last week, team member @cameron_pfiffer posted a snippet of code you could use to get structured LLM output from PDFs.

This week, we polished that code up a bit to showcase some of the cool things you can do with Outlines + PDFs.

A brief thread 👇

dottxt-ai.github.io/outlines/main/… @cameron_pfiffer The big trick here is to convert each page of a PDF to an image, and then use a vision language model like Qwen-2-VL or Pixtral. Most of your large providers do exactly this -- PDFs are among the worst machine-readable formats to ever exist, so treating them as images is simple.