nmhjklnm / llmflow

A lightweight framework for building LLM applications, with similar interface as LangChain, but more concise internal implementation.

English | 简体中文

LLMFlow

What's LLMFlow?

LLMFlow is a lightweight framework for building LLM applications. It offers an API interface similar to LangChain's LCEL, but with a more straightforward and simple internal implementation. Users familiar with LangChain can quickly get started with LLMFlow and extend or modify it as needed.

For example, LLMFlow implements the core functionality of LangChain's AgentExecutor in less than 300 lines of code. This is particularly beneficial for users who want to understand the underlying implementation of agents and make further optimizations.

LLMFlow is also very easy to use, as demonstrated by the following example of an Agent:

from core.tool import tool
from core.llm.openai.openai_llm import OpenAILLM
from core.agents.agent import Agent

llm = OpenAILLM(model="gpt-4o")


@tool
def multiply(left: int, right: int) -> int:
    """multiply"""
    return left * right


agent = Agent(llm=llm, tools=[multiply])
agent.invoke("what's 1024*2024?")

Why do we need LLMFlow?

Although LangChain provides a powerful framework for building LLM applications, its complex design also brings several issues with deeper use.

LangChain's extensive use of abstraction results in highly complex code, making it difficult to understand and maintain. While it initially simplifies the development process, as the complexity of requirements increases, these abstractions gradually become obstacles to productivity. Additionally, these abstractions increase the learning and debugging burden, forcing development teams to spend significant time dealing with internal framework code rather than focusing on application functionality.

Especially in today's rapidly evolving AI and LLM landscape, where new concepts and methods emerge weekly, a framework that is easy to understand and iterate on quickly is crucial!

To address these issues, LLMFlow adopts a more concise design with more refined code implementation, balancing abstraction and simplicity by the use of necessary abstractions without being overly complex. Ultimately, LLMFlow provides the same functionality and similar interfaces as LangChain but with significantly reduced code and fewer abstract concepts. This makes it easy to understand and modify LLMFlow, allowing for quick experimentation or customized development.

The project is still under development, with more features to be added. Your valuable feedback is welcome!