letsjson 0.1.5

Generate JSON that strictly matches a schema with automatic retries.

LetsJSON

Let LLMs generate exactly the JSON you define.

中文文档

Generate strongly constrained JSON from LLM outputs:

• Validate fields and types against your schema
• Auto-retry on invalid outputs (default: 3 attempts)
• Raise an error if all retries fail or return empty value
• It's very liteweight, with only 230+ lines of code.

Installation

uv add letsjson

or:

pip install letsjson

Usage

from openai import OpenAI
from letsjson import LetsJSON

client = OpenAI(
    api_key="your API key",
    base_url="your API base URL",
)
generator = LetsJSON(client, model="your model name")  # repeat is optional, default is 3

schema = {
    "title": str,
    "steps": [{"time": str, "location": str, "detail": str}],
}
result = generator.gen("Give me a 2-day London travel plan", schema)
print(result)

# Streaming output (optional)
result = generator.gen_stream(
    "Give me a 2-day London travel plan",
    schema,
    on_chunk=lambda chunk: print(chunk, end="", flush=True),
)
print("\n--- parsed json ---")
print(result)

# return:
# {
#   "title": "2-Day London Travel Plan",
#   "steps": [
#     {"time": "Day 1 Morning",
#      "location": "British Museum", 
#      "detail": "Explore ancient artifacts and world history."},
#    {"time": "Day 1 Afternoon",
#     "location": "Covent Garden",
#     "detail": "Enjoy street performances and shopping."},
#   ...
#  ]

Supported Schema Types

• Object: {"name": str, "age": int}
• List: {"items": [str]} (list schema must contain exactly one element type)
• Nested: {"user": {"name": str}, "tags": [str]}
• Strict type checks:
  • int does not accept bool
  • float accepts int and float (does not accept bool)