One post tagged with "LLM"

Reliable Structured Outputs with LLMs

March 1, 2025 · 3 min read

Curiosity brings awareness.

Ensuring Deterministic Outputs from LLMs

There are several strategies to obtain structured outputs from LLMs.

In Python, libraries such as Pydantic and Instructor facilitate structured output via JSON schema-based tool invocation. If you have the capability to host your own model, sglang is a viable option.

Pydantic validators are highly effective, provided that the input is in the form of a valid JSON string.

Let's see by example. For starters, here is the schema we want to parse.

from pydantic import BaseModel, ValidationError

class User(BaseModel):
    id: int
    name: str
    email: str
    active: bool = True  # default value

# JSON representation of the data
json_data = '''
{
    "id": 123,
    "name": "Alice",
    "email": "alice@example.com"
}
'''

try:
     # Directly validate and parse the JSON string
    user = User.model_validate_json(json_data)
    print("Validated Data:", user)
except ValidationError as e:
    print("Validation Error:", e.json())

This works. Pydantic has a pretty solid json to data model convertor. But it has to be a valid json string. Let's explore further.

# JSON representation of the data
# typical replies of a small LLM which does not adhere well to 'output_json' command
json_data = '''
Here is your json
{
    "id": 123,
    "name": "Alice",
    "email": "alice@example.com"
}
'''

try:
     # Directly validate and parse the JSON string using the new method
    user = User.model_validate_json(json_data)
    print("Validated Data:", user)
except ValidationError as e:
    print("Validation Error:", e.json())

Error is:

Validation Error: [{"type":"json_invalid","loc":[],"msg":"Invalid JSON: expected value at line 2 column 1","input":"\nHere is your json\n{\n    \"id\": 123,\n    \"name\": \"Alice\",\n    \"email\": \"alice@example.com\"\n}\n","ctx":{"error":"expected value at line 2 column 1"},"url":"https://errors.pydantic.dev/2.10/v/json_invalid"}]

Now, let's add one more step in the mix. Let's use the json_partial_py library to parse the JSON string. and then pass it to pydantic.

from json_partial_py import to_json_string # <---- this is a new import

# typical replies of a small LLM which does not adhere well to 'output_json' command
json_data = '''
Here is your json
{
    "id": 123,
    "name": "Alice",
    "email": "alice@example.com"
}
'''

try:
    stringified_json = to_json_string(json_data)
     # Directly validate and parse the JSON string using the new method
    user = User.model_validate_json(stringified_json)
    print("Validated Data:", user)
except ValidationError as e:
    print("Validation Error:", e.json())

and voila!! Now you can rest assured that you will get clean json parsed from the LLM output.

P.S. I am author of the json_partial_py library. It was extracted from baml project.

Ensuring Deterministic Outputs from LLMs​

Ensuring Deterministic Outputs from LLMs