llm_function

The llm_function library streamlines the creation of LLM prompt templates, schemas, validators, and fitness functions, making them reusable across various projects.

Experimental! Work in progress.

Installation

with pip:

pip install git+ssh://[email protected]/qwhex/llm_function.git

Providers

API Providers

Currently we support OpenAI and Mistral. You can install the necessary requirements this way:

pip install git+ssh://[email protected]/qwhex/llm_function.git[mistral]

or

pip install git+ssh://[email protected]/qwhex/llm_function.git[openai]

Llama.cpp Support

If you want to use it with llama.cpp, install the wrapper with the necessary install flags, e.g. for "cuda":

CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python

More info here

Usage

Define the prompt template (jinja2 + markdown)

Answer the following question as concisely as possible, while still including every important piece of information.
Include the source of the information if possible.
The answer shouldn't be longer than a few sentences.

Question:
"[[question]]"

Define the function (schema)

schema = {
    "args": {
        "question": v.str,
    },
    "response": {
        # you can define custom validators, e.g. "validate bash syntax"
        "validate": v.str
    }
}

# define the llm function
answer_concisely = create_llm_fn(
        template_path='answer_concisely.md',
        schema=schema,
        fitness=lambda resp: -1 * len(resp),  # the shorter, the better
)

# define the provider
provider = LlamaCppProvider(model_path='models/Phi-3-mini-4k-instruct-q4.gguf')

# use the function
questions = [
    "1+1=2. Why not 3?",
    "Why can't we tell whether someone has a consciousness?",
]
for question in questions:
    # print all generated answers, starting with the best one
    for response in answer_concisely({'question': question}, provider=provider, k=3):
        print(response)

JSON Schema Example

Create an appropriate prompt template. We recommend specifying a Typescript type for the expected response.

Generate a user profile for a fictional person.
Description: [[description]]

Follow the Typescript type below.

type UserProfile = {
id: string;
username: string;
age: number;
gender: "male" | "female" | "other";
interests: string[];
};

Answer with only JSON.

Define the JSON Schema under schema.response.json_schema When using the llama.cpp provider, the generation will be constrained to this schema. For all providers the response will be validated. You can define a custom validator which will be ran on the parsed, schema-validated response object.

schema = {
    "args": {
        "description": v.str
    },
    "response": {
        "json_schema": {
            "type": "object",
            "properties": {
                "id": {
                    "type": "string",
                    "format": "uuid"
                },
                "username": {
                    "type": "string",
                    "minLength": 3,
                    "maxLength": 20,
                    "pattern": "^[a-zA-Z0-9_]*$"
                },
                "age": {
                    "type": "integer",
                    "minimum": 18,
                    "maximum": 120
                },
                "gender": {
                    "type": "string",
                    "enum": ["male", "female", "other"]
                },
                "interests": {
                    "type": "array",
                    "items": {"type": "string"}
                },
            },
            "required": ["id", "username", "age", "gender"]
        },
        # you can define a custom validator for the parsed JSON below:
        "validator": some_custom_validator
    }
}

# define the llm function
generate_user_profile = create_llm_fn(
        template_path='user_profile.md',
        schema=schema,
)

Authors

All Rights Reserved
Copyright (c) 2024 Mice Pápai