llm-output-guard

Getting Started

Installation

llm-output-guard has no hard runtime dependencies. Install only what you need:

# Core only
pip install llm-output-guard

# Core + Pydantic v2
pip install "llm-output-guard[pydantic]"

# Core + JSON Schema validation
pip install "llm-output-guard[jsonschema]"

# Core + OpenAI integration
pip install "llm-output-guard[openai]"

# Core + LangChain integration
pip install "llm-output-guard[langchain]"

# Core + FastAPI integration
pip install "llm-output-guard[fastapi]"

# Everything
pip install "llm-output-guard[all]"

---

Core Concepts

Validator

Validator is the primary entry-point. You provide:

1. A schema — Pydantic model class, JSON Schema dict, or plain dict.
2. An LLM callable — any (prompt: str, **kwargs) -> str function.

from llm_output_guard import Validator

validator = Validator(
    schema=my_schema,
    llm_callable=my_llm,
    max_retries=2,              # default: 2
    retry_strategy="exponential",  # "fixed" | "exponential" | "linear"
    retry_delay=1.0,            # base delay in seconds
    raise_on_failure=False,     # raise MaxRetriesExceededError on failure?
    strict_json=False,          # require the full response to be valid JSON?
    system_prompt=None,         # optional system-level instructions
)

guard()

result = validator.guard("Your prompt here.")

• Augments the prompt with schema instructions.
• Calls the LLM.
• Extracts and parses JSON from the response.
• Validates against the schema.
• Retries on failure up to max_retries times.

validate_output()

Validate an existing string without calling the LLM:

result = Validator(schema=my_schema).validate_output('{"name": "Alice", "age": 30}')

GuardResult

Both methods return a GuardResult:

result.success       # bool
result.data          # validated object (Pydantic instance or dict), or None
result.raw_output    # raw LLM string
result.errors        # list of validation error dicts
result.attempts      # int — number of LLM calls made
result.schema_type   # "pydantic" | "json_schema" | "dict"

result.raise_for_status()   # raises ValidationError if not success
result.error_summary        # formatted string of all errors

---

Schema Types

Pydantic Model

from pydantic import BaseModel
from llm_output_guard import Validator

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

result = Validator(schema=Person, llm_callable=llm).guard("Who is Alice?")
person: Person = result.data   # fully typed

JSON Schema Dict

schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age":  {"type": "integer"},
    },
    "required": ["name", "age"],
}

result = Validator(schema=schema, llm_callable=llm).guard("Who is Alice?")
assert result.data["name"] == "Alice"

Plain Dict Schema

schema = {"name": str, "age": int}   # type-checks only

# Use Ellipsis to mark a field as required (any type)
strict_schema = {"name": ..., "score": float}

---

Retry Strategies

Strategy	Delay formula
fixed	base_delay
exponential	base_delay * 2^(attempt-1), with optional jitter
linear	base_delay * attempt

Validator(schema=schema, llm_callable=llm, retry_strategy="exponential", retry_delay=0.5)

---

CLI

# Validate output.json against schema.json
llm-guard validate output.json schema.json

# Print JSON Schema for a Pydantic model
llm-guard schema mypackage.models.Person --indent 4

This site is open source. Improve this page.