"Taming LLMs with Structured Outputs"
summary
Claims that guaranteed schema enforcement makes LLM outputs deterministic by constraining generation to a grammar derived from a JSON schema and shows it working with Zod + ChatGPT API (OpenAI feature introduced this summer). Demos include parsing form fields reliably, using function calling to fetch live weather, and converting course materials into a 30-question Jeopardy (6x5) game. Code examples run (Node.js/npm) and produce valid JSON every time rather than brittle freeform text.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:47LLMs consistency struggles
- 3:31LLMs without safeguards
- 4:56Prompt engineering limits
- 6:42Structured outputs enforce schemas
- 8:26ChatGPT API implementation
- 11:19Extracting structured info
- 12:43Function calling use case
- 14:51Real-world Ziggy platform
- 16:18AI future and closing
related talks
47mStructured Outputs in OpenAI: Building Reliable JSON-Based AI Integrations
13mGetting LLMs to do what you want: Output controllers
26mAgentic Fundamentals: How language models and code can communicate
29mKeynote: LLMs, what are they good for?
35mBuilding LLM-based Agents
01h 00mFrom LLM Theory to Practical Agentic Implementations
47mHarnessing LLMs for Next-Level Security Analysis
01h 15m10x Development: LLMs For the working Programmer