LLM-Client-SDK

LLM-Client-SDK is an SDK for seamless integration with generative AI large language models (We currently support - OpenAI, Google, AI21, HuggingfaceHub, Aleph Alpha, Anthropic, Local models with transformers - and many more soon).

Our vision is to provide async native and production ready SDK while creating a powerful and fast integration with different LLM without letting the user lose any flexibility (API params, endpoints etc.). *We also provide sync version, see more details below in Usage section.

Base Interface

The package exposes two simple interfaces for seamless integration with LLMs (In the future, we will expand the interface to support more tasks like list models, edits, etc.):

from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import Any, Optional
from enum import Enum
from dataclasses_json import dataclass_json, config
from aiohttp import ClientSession


class BaseLLMClient(ABC):
    @abstractmethod
    async def text_completion(self, prompt: str, **kwargs) -> list[str]:
        raise NotImplementedError()

    async def get_tokens_count(self, text: str, **kwargs) -> int:
        raise NotImplementedError()


class Role(Enum):
    SYSTEM = "system"
    USER = "user"
    ASSISTANT = "assistant"


@dataclass_json
@dataclass
class ChatMessage:
    role: Role = field(metadata=config(encoder=lambda role: role.value, decoder=Role))
    content: str
    name: Optional[str] = field(default=None, metadata=config(exclude=lambda name: name is None))
    example: bool = field(default=False, metadata=config(exclude=lambda _: True))
    

@dataclass
class LLMAPIClientConfig:
    api_key: str
    session: ClientSession
    base_url: Optional[str] = None
    default_model: Optional[str] = None
    headers: dict[str, Any] = field(default_factory=dict)


class BaseLLMAPIClient(BaseLLMClient, ABC):
    def __init__(self, config: LLMAPIClientConfig):
        ...

    @abstractmethod
    async def text_completion(self, prompt: str, model: Optional[str] = None, max_tokens: int | None = None,
                              temperature: Optional[float] = None, top_p: Optional[float] = None, **kwargs) -> list[str]:
        raise NotImplementedError()

    async def chat_completion(self, messages: list[ChatMessage], temperature: float = 0,
                              max_tokens: int = 16, model: Optional[str] = None, **kwargs) -> list[str]:
        raise NotImplementedError()

    async def embedding(self, text: str, model: Optional[str] = None, **kwargs) -> list[float]:
        raise NotImplementedError()

    async def get_chat_tokens_count(self, messages: list[ChatMessage], **kwargs) -> int:
        raise NotImplementedError()

Requirements

Python 3.9+

Installation

If you are worried about the size of the package you can install only the clients you need, by default we install none of the clients.

For all current clients support

$ pip install llm-client[all]

For only the base interface and some light LLMs clients (AI21 and Aleph Alpha)

$ pip install llm-client

Optional Dependencies

For all current api clients support

$ pip install llm-client[api]

For only local client support

$ pip install llm-client[local]

For sync support

$ pip install llm-client[sync]

For only OpenAI support

$ pip install llm-client[openai]

For only HuggingFace support

$ pip install llm-client[huggingface]

Usage

Using OpenAI directly through OpenAIClient - Maximum control and best practice in production

import os
from aiohttp import ClientSession
from llm_client import ChatMessage, Role, OpenAIClient, LLMAPIClientConfig

OPENAI_API_KEY = os.environ["API_KEY"]
OPENAI_ORG_ID = os.getenv("ORG_ID")


async def main():
    async with ClientSession() as session:
        llm_client = OpenAIClient(LLMAPIClientConfig(OPENAI_API_KEY, session, default_model="text-davinci-003",
                                                     headers={"OpenAI-Organization": OPENAI_ORG_ID}))  # The headers are optional
        text = "This is indeed a test"
        messages = [ChatMessage(role=Role.USER, content="Hello!"),
                    ChatMessage(role=Role.SYSTEM, content="Hi there! How can I assist you today?")]

        print("number of tokens:", await llm_client.get_tokens_count(text))  # 5
        print("number of tokens for chat completion:", await llm_client.get_chat_tokens_count(messages, model="gpt-3.5-turbo"))  # 23
        print("generated chat:", await llm_client.chat_completion(messages, model="gpt-3.5-turbo"))  # ['Hi there! How can I assist you today?']
        print("generated text:", await llm_client.text_completion(text))  # [' string\n\nYes, this is a test string. Test strings are used to']
        print("generated embedding:", await llm_client.embedding(text))  # [0.0023064255, -0.009327292, ...]

Using LLMAPIClientFactory - Perfect if you want to move fast and to not handle the client session yourself

import os
from llm_client import LLMAPIClientFactory, LLMAPIClientType

OPENAI_API_KEY = os.environ["API_KEY"]


async def main():
    async with LLMAPIClientFactory() as llm_api_client_factory:
        llm_client = llm_api_client_factory.get_llm_api_client(LLMAPIClientType.OPEN_AI,
                                                               api_key=OPENAI_API_KEY,
                                                               default_model="text-davinci-003")

        await llm_client.text_completion(prompt="This is indeed a test")
        await llm_client.text_completion(prompt="This is indeed a test", max_tokens=50)

        
# Or if you don't want to use async
from llm_client import init_sync_llm_api_client

llm_client = init_sync_llm_api_client(LLMAPIClientType.OPEN_AI, api_key=OPENAI_API_KEY,
                                      default_model="text-davinci-003")

llm_client.text_completion(prompt="This is indeed a test")
llm_client.text_completion(prompt="This is indeed a test", max_tokens=50)

Local model

import os
from transformers import AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoTokenizer
from llm_client import LocalClientConfig, LocalClient

async def main():
    try:
        model = AutoModelForCausalLM.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
    except ValueError:
        model = AutoModelForSeq2SeqLM.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
    tokenizer = AutoTokenizer.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
    llm_client = LocalClient(LocalClientConfig(model, tokenizer, os.environ["TENSORS_TYPE"], os.environ["DEVICE"]))

    await llm_client.text_completion(prompt="This is indeed a test")
    await llm_client.text_completion(prompt="This is indeed a test", max_tokens=50)


# Or if you don't want to use async
import async_to_sync

try:
    model = AutoModelForCausalLM.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
except ValueError:
    model = AutoModelForSeq2SeqLM.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
tokenizer = AutoTokenizer.from_pretrained(os.environ["MODEL_NAME_OR_PATH"])
llm_client = LocalClient(LocalClientConfig(model, tokenizer, os.environ["TENSORS_TYPE"], os.environ["DEVICE"]))

llm_client = async_to_sync.methods(llm_client)

llm_client.text_completion(prompt="This is indeed a test")
llm_client.text_completion(prompt="This is indeed a test", max_tokens=50)

Contributing

Contributions are welcome! Please check out the todos below, and feel free to open issue or a pull request.

Todo

The list is unordered

• Add support for more LLMs
  • Anthropic
  • Google
  • Cohere
• Add support for more functions via LLMs
  • embeddings
  • chat
  • list models
  • edits
  • more
• Add contributing guidelines and linter
• Create an easy way to run multiple LLMs in parallel with the same prompts
• Convert common models parameter
  • temperature
  • max_tokens
  • top_p
  • more

Development

To install the package in development mode, run the following command:

$ pip install -e ".[all,test]"

To run the tests, run the following command:

$ pytest tests

If you want to add a new LLMClient you need to implement BaseLLMClient or BaseLLMAPIClient.

If you are adding a BaseLLMAPIClient you also need to add him in LLMAPIClientFactory.

You can add dependencies to your LLMClient in pyproject.toml also make sure you are adding a matrix.flavor in test.yml.