🔗 🐍 Lang Interface

Super lightweight helper to turn your python interface into an AI assistant.

pip install lang-interface

🚀 Quick start

from dotenv import load_dotenv
from os import environ
from openai import OpenAI

from lang_interface import Assistant
environ['OPENAI_API_KEY'] = '<api_key>'


class MyAPI:
    """Api for managing user's list of contacts (mobile numbers)"""
    contacts = {'John': '000-000', 'Bill': '111-111'}

    def do_get_contact_list(self, name_starts_with: str = None) -> dict[str, str]:
        """Get contacts names and phones"""
        return {
            name: phone
            for name, phone in self.contacts.items()
            if name.startswith(name_starts_with)
        }

    def do_add_contact(self, name: str, phone: str) -> str:
        """Add new contact"""
        if name in self.contacts:
            raise Exception(f'Contact with name {name} already exists!')
        self.contacts[name] = phone
        

llm = OpenAI()
api = MyAPI()
assistant = Assistant(api, llm)
print(assistant('Do I have Bob in my contacts?'))

Example interactive mode 💬

def example_chat():
    while True:
        try:
            q = input('\033[1;36m> ')
            print('\033[0m', end='')
            answer = assistant(q)
            print(f'\033[0massistant: {answer}')
        except KeyboardInterrupt:
            print('\033[0;32mBuy!')


example_chat()

📝 Basics

Lang Interface uses python docstrings and type hints to create a short specification of the programming API for LLM.

The quality of outputs depends on well-structured class, where docstrings are laconic and not ambiguous. It is recommended to use python typing hits to describe parameters and return values. If you need to specify complicated input/output format use Pydantic models:

from pydantic import BaseModel

class MyContact(BaseModel):
    id: int
    name: str
    phone_number: str
    created_at: datetime

class Interface:
    def do_create_contact(self, contact: MyContact):
        ...

However, using dictionaries would still be more reliable, but remember to write a comprehensible docstring.

LLM

lang_interface supports OpenAI client or any callable object:

def call_llm(messages: list[dict]) -> str: ...

🔒 Security concerns

Giving the API in hands of llm make sure you have all safety checks to prevent from malicious actions being made by llm generated instructions. Take a look at this example:

import os
assistant = Assistant(os, openai_client)

In this example the whole os module is given as an API to LLM, potentially making it possible to call rm -rf / even I LLM was never asked to do so. Providing an API make sure LLM cannot harm your data or system in any way.

💻️ Advanced

Classes vs Modules

lang_interface supports both: python module and a class as an API handler. For example:

"""My module for managing ..."""
def do_this():...
def do_that():...

Or:

class MyAPI:
    """My API class for managing ..."""
    def do_this(self):...
    def do_that(self):...

Prefixes

By default lang_interface scans all public methods/functions. If you need to specify specific set of methods, use methods_prefix:

Assistant(api, llm, methods_prefix='do_')

Debug

Use DEBUG=True, to print all interactions will LLM

import lang_interface
lang_interface.DEBUG = True

Callbacks

You might need to get a callback on every LLM request/response. You can do that providing a custom callable object as an LLM:

class MyLLM:
    def __call__(self, messages: list[dict]) -> str:
        resp = openai_client.chat.completions.create(
            messages=messages, model='gpt-4o'
        )
        text = resp.choices[0].message.content
        logger.info(f"Response from LLM: {text}")
        
        return text