Implementation:Sgl project Sglang Json Output Parsing

Knowledge Sources	Pydantic Documentation SGLang
Domains	NLP, Data_Validation, Structured_Generation
Last Updated	2026-02-10 00:00 GMT

Overview

Wrapper documentation for parsing and validating JSON output from constrained generation using standard library json and Pydantic.

Description

This covers two external APIs used together: json.loads() from the Python standard library for basic JSON parsing, and BaseModel.model_validate_json() from Pydantic for type-safe validation. When used after SGLang's constrained generation, syntactic validity is already guaranteed, so these calls primarily provide type conversion and semantic validation.

Usage

Call json.loads() on the generated text to get a Python dict. For type-safe objects, call YourModel.model_validate_json() to get a validated Pydantic model instance.

Code Reference

Source Location

• Library: Python stdlib json module
• Library: Pydantic (external)
• Usage pattern: examples/frontend_language/usage/json_decode.py

Signature

# Standard library
import json
result_dict = json.loads(generated_text: str) -> dict

# Pydantic validation
from pydantic import BaseModel
result_model = YourModel.model_validate_json(generated_text: str) -> YourModel

Import

import json
from pydantic import BaseModel

I/O Contract

Inputs

Name	Type	Required	Description
generated_text	str	Yes	JSON text from constrained generation

Outputs

Name	Type	Description
result_dict	dict	Parsed Python dictionary (via json.loads)
result_model	BaseModel	Validated Pydantic model instance (via model_validate_json)

Usage Examples

Basic JSON Parsing

import json

output = engine.generate(
    "Generate person info as JSON:",
    {"regex": regex_from_schema, "max_new_tokens": 128},
)

# Parse to dict
data = json.loads(output["text"])
print(data["name"], data["age"])

Pydantic Validation

from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int
    email: str

output = engine.generate(prompt, {"regex": regex, "max_new_tokens": 128})

# Validate with type checking
person = Person.model_validate_json(output["text"])
print(person.name)   # Typed str
print(person.age)    # Typed int

External Reference

• Pydantic JSON Documentation

Related Pages

Implements Principle

• Principle:Sgl_project_Sglang_Structured_Output_Validation