Ask HN: How does structured output from LLMs work under the hood?

2 dcreater 0 5/7/2025, 5:48:25 PM

OpenAI, Ollama, LiteLLM python packages allow you to set the response format to a Pydantic model. As I understand that this just, serializes the model into a JSON. Is the JSON then just passed as context to the LLM and the LLM asked to adhere to the provided JSON? Or is there something more technical/deterministic happening that constrains the output to the JSON format provided?

Redis is open source again (antirez.com)

I'd rather read the prompt (claytonwramsey.com)

The curse of knowing how, or; fixing everything (notashelf.dev)

Show HN: Clippy – 90s UI for local LLMs (felixrieseberg.github.io)

Judge rules Apple executive lied under oath, makes criminal contempt referral (thebignewsletter.com)

Ty: A fast Python type checker and language server (github.com)

Design for 3D-Printing (blog.rahix.de)

Technical analysis of the Signal clone used by Trump officials (micahflee.com)

Show HN: Free, in-browser PDF editor (breezepdf.com)

Claude Integrations (anthropic.com)

Gemini 2.5 Pro Preview (developers.googleblog.com)

The Death of Daydreaming (afterbabel.com)

Zed: High-performance AI Code Editor (zed.dev)

OpenAI reaches agreement to buy Windsurf for $3B (bloomberg.com)

Starting July 1, academic publishers can't paywall NIH-funded research (nih.gov)

Accountability Sinks (250bpm.substack.com)

Claude's system prompt is over 24k tokens with tools (github.com)

Evolving OpenAI's Structure (openai.com)

Matt Godbolt sold me on Rust by showing me C++ (collabora.com)

CLion Is Now Free for Non-Commercial Use (blog.jetbrains.com)

How to live an intellectually rich life (utsavmamoria.substack.com)

I decided to pay off a school’s lunch debt (huffpost.com)

Waiting for Postgres 18: Accelerating Disk Reads with Asynchronous I/O (pganalyze.com)

We fell out of love with Next.js and back in love with Ruby on Rails (hardcover.app)

Show HN: I built a synthesizer based on 3D physics (anukari.com)

Show HN: Real-time AI Voice Chat at ~500ms Latency (github.com)

Unity’s Open-Source Double Standard: the ban of VLC (mfkl.github.io)

Why Archers Didn't Volley Fire (acoup.blog)

A faster way to copy SQLite databases between computers (alexwlchan.net)

DuckDB is probably the most important geospatial software of the last decade (dbreunig.com)

India launches attack on 9 sites in Pakistan and Pakistani Jammu and Kashmir (reuters.com)

Mistral ships Le Chat – enterprise AI assistant that can run on prem (mistral.ai)

Apple App Store guidelines remove ban on encouraging external payments in US (developer.apple.com)

Buffett to step down following six-decade run atop Berkshire (bloomberg.com)

Gorgeous-GRUB: collection of decent community-made GRUB themes (github.com)

Third party cookies must be removed (w3ctag.github.io)

Time saved by AI offset by new work created, study suggests (arstechnica.com)

Curl: We still have not seen a valid security report done with AI help (linkedin.com)

Replacing Kubernetes with systemd (2024) (blog.yaakov.online)

VVVVVV Source Code (github.com)

The vocal effects of Daft Punk (bjango.com)

So Much Blood (dynomight.net)

The language brain matters more for programming than the math brain? (2020) (massivesci.com)

Old Soviet Venus descent craft nearing Earth reentry (leonarddavid.com)

Sneakers (1992) – 4K makeover sourced from the original camera negative (blu-ray.com)

“Fewer Users” Warning Hurting Specialized and New Apps (support.google.com)

International Workers' Day (en.wikipedia.org)

An appeal to Apple from Anukari (anukari.com)

Launch HN: Exa (YC S21) – The web as a database

Judge said Meta illegally used books to build its AI (wired.com)

Ask HN: How does structured output from LLMs work under the hood?

Comments (0)