← BACK TO AI LAB

AI LAB / LLMs / Structured Output

Can LLMs Generate Schemas?

Measuring structured schema generation across models, prompts, and repair strategies.

LLMsStructured OutputEvaluation
EXPERIMENT
EXP-016
STATUS
complete
DATE
Aug 18, 2026
TAGS
LLMs · Structured Output · Evaluation

EXPERIMENT / HYPOTHESIS

Schema-constrained generation plus one validation-aware repair will exceed 95% syntactic validity.

  1. 01Generate 80 schemas from product descriptions
  2. 02Validate syntax and semantic constraints
  3. 03Retry invalid output once with structured errors
First-pass valid
89.8%
After repair
97.5%
Runs
80

CONCLUSIONA single targeted repair recovered most syntax failures, but semantic field mistakes still required domain validation.

01

Test design

Prompts ranged from simple forms to nested workflow contracts with discriminated variants and cross-field constraints.

  • Syntax validity
  • Constraint coverage
  • Unnecessary fields
02

Finding

Provider schema modes solved formatting but not product semantics. Deterministic domain checks remained necessary.

  • Validate twice
  • Repair narrowly
  • Do not trust plausible fields