OpenAI JSON mode, structured outputs, grammar-constrained generation.

JSON Mode — Structured Outputs from Language Models

JSON mode forces an LLM to emit valid JSON output. OpenAI JSON mode guarantees syntactically valid JSON. OpenAI Structured Outputs (2024) goes further — guarantees the response matches a specific JSON Schema. Anthropic (extract JSON from text), Grammar-constrained generation (llama.cpp, outlines library) enable local structured outputs. Structured outputs eliminate brittle string parsing and JSON repair hacks from production code.

Interactive Pipeline Flow

UnderstandImplementOptimise

📐 Schema — define JSON shape

Define the JSON schema: what fields, what types, what constraints. This is both documentation and enforcement.

Architecture Diagrams

Full pipeline: OpenAI structured output + Zod validation.

Loading diagram...

Key Design Decisions

📐Structured Outputs eliminate JSON parsing bugs entirely

OpenAI Structured Outputs with strict: true guarantees the response matches your JSON Schema. No JSON.parse failures, no repair hacks, no schema mismatch surprises. Use it whenever available.

✅Always validate parsed JSON with Zod even with Structured Outputs

Even with strict mode, numeric fields might be integers when you expect floats, or arrays might be empty when you expect at least one item. Zod catches type-level issues that JSON Schema validations may miss.

🔷z.infer gives you free TypeScript types from Zod schema

Define the Zod schema once: const S = z.object({...}). Get the JSON schema: zodToJsonSchema(S). Get the TypeScript type: z.infer<typeof S>. One definition, three uses, zero duplication.

Discussion