Introduction

Gentopia is a lightweight and extensible framework for LLM driven autonomous Agent. It provides multiple essential components to build, test and evaluate agents. At its core, Gentopia aims to embody agents with single config files, thus minimizing your effort in maintaining, tuning, and sharing agents.

Gentopia maintains a pool (GentPool) to share public agents specialized for different tasks. In this platform, you could easily call other agents for interaction, extend and optimize agent components towards your needs. You can also publish your specialized agent to get our unique benchmark eval and share your agent with the world.

Motivation

Think of building a real AI assistant. Do we really need a >> 175B GPT-4 to summarize Web search or print("helloworld")? Recent studies like TinyStories, Specializing Reasoning, Let’s Verify SbS, ReWOO, etc. point us towards an intuitive yet undervalued direction 👉

An LLM is more capable if you create a context/distribution shift narrowed to some target tasks.

Sadly, there is no silver bullet for agent specialization. For example, you can

  • Simply add Let's think step by step. in your prompt for more accurate Math QA.

  • Give a few-shot exemplar in your prompt to guide a better reasoning trajectory for novel plotting.

  • Supervise fine-tuning (SFT) your 70B llama2 like this to match reasoning of 175B GPT-3.5.

  • Tune your agent paradigm like this demo to easily half the execution time for Seach & Summarize.

  • And more …

Isn’t that beautiful if one shares his specialized intelligence, allowing others to reproduce, build on, or interact with it at ease? This belief inspires us to build Gentopia, a convenient framework for agent specialization, sharing, and interaction, encouraging collective growth to greater intelligence.

Gentopia

Gentopia is the core framework, providing pre-built and extensible components to assemble agents.

Gentopia Architecture

We provide an expanding number of essential components – Agent type, LLM, Plugins, Memory, etc. We also make the base classes simple and lightweight, so you can easily extend and customize your own components.

GentPool

GentPool is an application built on Gentopia, providing a platform for agent sharing, interaction and evaluation. It consists of two parts:

  • pool: The agent zoo. Each agent consists of a core config file and some optional companion files defining custom components.

  • bench: Our unique agent eval benchmark. It tests a wide range of agent capability beyond vanilla LLMs.

Gentbench Architecture