task_categories:
- question-answering
- translation
- summarization
- text-generation
- text2text-generation
- conversational
tags:
- agent
- multi-agent
- autogpt
- autogen
- agentgpt
- gptq
- wizard
- code-generation
- retrieval-augmented-generation
- humaneval
Roy: Rapid Prototyping of Agents with Hotswappable Components
Roy is a lightweight alternative to autogen for developing advanced multi-agent systems using language models. It aims to simplify and democratize the development of emergent collective intelligence.
Features
Model Agnostic: Use any LLM, no external APIs required. Defaults to a 4-bit quantized wizard-coder-python model for efficiency.
Modular and Composable: Roy decomposes agent interactions into reusable building blocks - templating, retrieving, generating, executing.
Transparent and Customizable: Every method has a clear purpose. Easily swap out components or add new capabilities.
Quickstart
git clone https://github.com/JosefAlbers/Roy
cd Roy
pip install -r requirements.txt
pip install -U transformers optimum accelerate auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
from roy import Roy, Roys
roy = Roy()
s = '"What date is today? Which big tech stock has the largest year-to-date gain this year? How much is the gain?'
roy.generate(roy.format(s))
Rapid Benchmarking
Roy provides a simple way to evaluate and iterate on your model architecture.. This allows you to:
Easily swap out components, such as language models, prompt formats, agent architectures, etc
Benchmark on different tasks like arithmetic, python coding, etc (default is OpenAI's HumanEval)
Identify agent's areas of strengths and weaknesses
from Roy.util import piecewise_human_eval
# Comparing different language models
piecewise_human_eval(0, lm_id='TheBloke/WizardCoder-Python-7B-V1.0-GPTQ')
# -> {'pass@1': 0.6341463414634146}
piecewise_human_eval(0, lm_id='TheBloke/tora-code-7B-v1.0-GPTQ')
# -> {'pass@1': 0.5609756097560976}
piecewise_human_eval(0, lm_id='TheBloke/Arithmo-Mistral-7B-GPTQ')
# -> {'pass@1': 0.5121951219512195}
# Testing a custom agent architecture
piecewise_human_eval(0, fx=<your_custom_Roy_agent>)
Takes around 30 minutes each on a free Google Colab runtime.
Constrained Beam Search
Use templates to structure conversations (control output length, format, etc)
roy.generate(s, ('\n```python', '\n```')) # Generate a python code block
roy.generate(s, (('\n```python', '\n```javascript'), '\n```')) # Generate python or javascript codes
roy.generate(s, ('\n```python', 100, '\n```')) # Generate a code block of size less than 100 tokens
Retrieval Augmented Generation
Enhance generation with relevant knowledge.
s = 'Create a text to image generator.'
r = roy.retrieve(s, n_topk=3, src='huggingface')
[roy.generate(s) for s in r]
Auto-Feedback
Agents recursively improve via critiquing each other.
s = "Create a secure and unique secret code word with a Python script that involves multiple steps to ensure the highest level of confidentiality and protection.\n"
for i in range(2):
c = roy.generate(s, prohibitions=['input'])
s += roy.execute(c)
Auto-Grinding
Agents collaborate in tight loops to iteratively refine outputs to specification.
user_request = "Compare the year-to-date gain for META and TESLA."
ai_response = roy.generate(user_request, ('\n```python', ' yfinance', '\n```'))
for i in range(2):
shell_execution = roy.execute(ai_response)
if 'ModuleNotFoundError' in shell_execution:
roy.execute(roy.generate(roy.format(f'Write a shell command to address the error encountered while running this Python code:\n\n{shell_execution}')))
elif 'Error' in shell_execution:
ai_response = roy.generate(roy.format(f'Modify the code to address the error encountered:\n\n{shell_execution}'))
else:
break
Multi-Agent
Flexible primitives to build ecosystems of agents.
roys = Roys()
# AutoFeedback
roys.create(agents = {'Coder': 'i = execute(generate(i))'})
roys.start(requests = {'i': 'Create a mobile application that can track the health of elderly people living alone in rural areas.'})
# Retrieval Augmented Generation
roys.create(
agents = {
'Retriever': 'r = retrieve(i)',
'Generator': 'o = generate(r)',
})
roys.start(requests = {'i': 'Create a Deutsch to English translator.'})
# Providing a custom tool to one of the agents using lambda
roys.create(
agents = {
'Coder': 'c = generate(i)',
'Proxy': 'c = custom(execute(c))',
},
tools = {'custom': lambda x:f'Modify the code to address the error encountered:\n\n{x}' if 'Error' in x else None})
roys.start(requests = {'i': 'Compare the year-to-date gain for META and TESLA.'})
# Another way to create a custom tool for agents
def custom_switch(self, c):
py_str = 'Modify the code to address the error encountered:\n\n'
sh_str = 'Write a shell command to address the error encountered while running this Python code:\n\n'
x = self.execute(c)
if 'ModuleNotFoundError' in x:
self.execute(self.generate(sh_str+x))
elif 'Error' in x:
self.dict_cache['i'] = [py_str+x]
else:
return '<<<Success>>>:\n\n'+x
roys.create(
agents = {
'Coder': 'c = generate(i)',
'Proxy': '_ = protocol(c)',
},
tools = {'protocol': custom_switch})
roys.start(requests = {'i': 'Compare the year-to-date gain for META and TESLA.'})
Emergent Multi-Agent Dynamics
Roy aims to facilitate the emergence of complex, adaptive multi-agent systems. It draws inspiration from biological and AI concepts to enable decentralized coordination and continual learning.
Survival of the Fittest - Periodically evaluate and selectively retain high-performing agents based on accuracy, speed etc. Agents adapt through peer interactions.
Mixture of Experts - Designate agent expertise, dynamically assemble specialist teams, and route tasks to optimal experts. Continuously refine and augment experts.
These mechanisms facilitate the emergence of capable, adaptive, and efficient agent collectives.
Get Involved
Roy is under active development. We welcome contributions - feel free to open issues and PRs!
Support the Project
If you found this project helpful or interesting and want to support more of these experiments, feel free to buy me a coffee!
