Instructor — Typed Structured Outputs for LLMs

Practical Notes

Use Instructor when you already know the shape of the answer and you want the model to fill it in with high reliability. Start with one Pydantic model per call (e.g., User, ProductReview), add tight field constraints (enums, ranges), then layer in retry budgets. Once stable, treat the schema as an API contract: version it, add regression examples, and monitor validation failures as a quality metric.

Safety note: Schema discipline matters: oversized models and ambiguous fields cause retries, latency, and cost spikes.

FAQ

Q: What problem does Instructor solve? A: It enforces a typed schema on LLM outputs (Pydantic) and retries invalid generations, so you ship structured results instead of brittle parsing.

Q: Is it only for OpenAI? A: No. The repo shows provider strings for OpenAI, Anthropic, Google, and local Ollama; the API stays consistent.

Q: How do I reduce failures? A: Keep the schema minimal, constrain enums, and ask for one object per call; smaller schemas validate more reliably and retry less.