AI-UI-GmbH / SchemaLLM

An intuitive and robust way to generate structured text. Inspired by https://github.com/1rgs/jsonformer

Schemallm

An intuitive and robust way to generate structured text. Inspired by Jsonformer

Base functions

supported schema types:

• number
• boolean
• string
• array
• object

What's new

• Pydantic is supported
• Literal is supported
• Tuple is supported

Usage

# define schema
from typing import Tuple
from pydantic import BaseModel, Field

class Employer(BaseModel):
    name: str
    location: str

class Person(BaseModel):
    name: str
    age: int
    employer: Employer
    job: str
    family_members: Tuple[str, str] 
    # or
    # family_members: List[str] = Field(minItems=2, maxItems=2)

# load model
from schemallm import SchemaLLM
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("TheBloke/Llama-2-7B-Chat-GPTQ")
tokenizer = AutoTokenizer.from_pretrained("TheBloke/Llama-2-7B-Chat-GPTQ")
schema_llm = SchemaLLM(
    model=model,
    tokenizer=tokenizer
)
output = schema_llm.generate(prompt=input_text, schema=schema)
print(output)
# {'name': 'Alice', 'age': 21, 'employer': {'name': 'LMQL Inc', 'location': 'Zurich'}, 'job': 'engineer', 'family_members': ['husband', 'son']}

other examples:

• calling function with arguments from local llm

Install

pip install git+https://github.com/pfan21/schemallm