llm-output-guard

API Reference

llm_output_guard.Validator

The main class for guarding LLM outputs.

class Validator(
    schema: SchemaDefinition,
    llm_callable: Optional[LLMCallable] = None,
    *,
    max_retries: int = 2,
    retry_strategy: str = "exponential",
    retry_delay: float = 1.0,
    raise_on_failure: bool = False,
    strict_json: bool = False,
    system_prompt: Optional[str] = None,
)

Parameters

Parameter	Type	Default	Description
schema	Pydantic class / dict	—	Schema to validate against
llm_callable	(str, **kwargs) -> str	None	Callable that calls the LLM
max_retries	int	2	Extra retry attempts on failure
retry_strategy	str	"exponential"	"fixed" / "exponential" / "linear"
retry_delay	float	1.0	Base delay (seconds) between retries
raise_on_failure	bool	False	Raise MaxRetriesExceededError if all retries fail
strict_json	bool	False	Require entire response to be valid JSON
system_prompt	str	None	System-level instructions prepended to every prompt

Methods

guard(prompt, **llm_kwargs) -> GuardResult

Call the LLM and validate. Retries automatically on failure.

validate_output(raw_output) -> GuardResult

Validate an existing string. No LLM call is made.

---

llm_output_guard.GuardResult

@dataclass
class GuardResult:
    success: bool
    data: Optional[Any]
    raw_output: str
    errors: List[Dict[str, Any]]
    attempts: int
    schema_type: str

Methods

raise_for_status() -> GuardResult

Raises ValidationError if success is False. Returns self otherwise.

error_summary -> str (property)

Human-readable summary of all validation errors.

---

llm_output_guard.SchemaParser

Low-level schema detection and validation.

parser = SchemaParser(schema)
parser.schema_type        # SchemaType enum
parser.json_schema        # JSON Schema dict representation
parser.describe()         # human-readable description string
result, errors = parser.validate(data)  # validate a dict/list

---

Exceptions

Exception	Description
LLMOutputGuardError	Base class for all errors
ValidationError	Schema validation failed
SchemaParseError	Schema cannot be parsed
JSONParseError	LLM output cannot be parsed as JSON
MaxRetriesExceededError	All retry attempts exhausted
IntegrationError	Third-party integration error

---

Integrations

llm_output_guard.integrations.openai.GuardedOpenAI

GuardedOpenAI(
    schema, *, model="gpt-4o-mini", api_key=None, temperature=0.0,
    max_tokens=None, max_retries=2, retry_strategy="exponential",
    raise_on_failure=False, system_prompt=None, extra_openai_kwargs=None
)

llm_output_guard.integrations.langchain.GuardedLLM

GuardedLLM(llm, schema, *, max_retries=2, retry_strategy="exponential", raise_on_failure=False)
result = guarded.invoke(prompt)

llm_output_guard.integrations.langchain.GuardOutputParser

parser = GuardOutputParser(schema)
data = parser.parse(text)            # raises ValidationError on failure
instructions = parser.get_format_instructions()

llm_output_guard.integrations.fastapi.guarded_endpoint

@app.post("/route")
@guarded_endpoint(schema=MyModel, llm_callable=my_llm, max_retries=2)
async def route(request: Request):
    return prompt_string

---

Utilities

llm_output_guard.utils.json_helpers

Function	Description
extract_json(text)	Extract JSON from prose/markdown
parse_json_safely(text)	Parse JSON, raise JSONParseError on failure
to_json_string(obj)	Serialise object to indented JSON

llm_output_guard.utils.validators

Function	Description
is_non_empty_string(v)	Non-empty string check
is_positive_number(v)	Positive int/float check
is_valid_email(v)	Basic email format check
is_valid_url(v)	Basic HTTP/HTTPS URL check
is_non_empty_list(v)	Non-empty list check
validate_fields(data, rules)	Apply validators to specific fields

This site is open source. Improve this page.