search

microsoft / TinyTroupe

LLM-powered multiagent persona simulation for imagination enhancement and business insights.
MIT License
4.14k stars 288 forks source link

readme

TinyTroupe 🤠🤓🥸🧐

LLM-powered multiagent persona simulation for imagination enhancement and business insights.

A tiny office with tiny people doing some tiny jobs.

TinyTroupe is an experimental Python library that allows the simulation of people with specific personalities, interests, and goals. These artificial agents - TinyPersons - can listen to us and one another, reply back, and go about their lives in simulated TinyWorld environments. This is achieved by leveraging the power of Large Language Models (LLMs), notably GPT-4, to generate realistic simulated behavior. This allow us to investigate a wide range of convincing interactions and consumer types, with highly customizable personas, under conditions of our choosing. The focus is thus on understanding human behavior and not on directly supporting it (like, say, AI assistants do) -- this results in, among other things, specialized mechanisms that make sense only in a simulation setting. Further, unlike other game-like LLM-based simulation approaches, TinyTroupe aims at enlightening productivity and business scenarios, thereby contributing to more successful projects and products. Here are some application ideas to enhance human imagination:

In all of the above, and many others, we hope experimenters can gain insights about their domain of interest, and thus make better decisions.

We are releasing TinyTroupe at a relativelly early stage, with considerable work still to be done, because we are looking for feedback and contributions to steer development in productive directions. We are particularly interested in finding new potential use cases, for instance in specific industries.

[!NOTE] 🚧 WORK IN PROGRESS: expect frequent changes. TinyTroupe is an ongoing research project, still under very significant development and requiring further tidying up. In particular, the API is still subject to frequent changes. Experimenting with API variations is essential to shape it correctly, but we are working to stabilize it and provide a more consistent and friendly experience over time. We appreciate your patience and feedback as we continue to improve the library.

[!CAUTION] ⚖️ Read the LEGAL DISCLAIMER. TinyTroupe is for research and simulation only. You are fully responsible for any use you make of the generated outputs. Various important additional legal considerations apply and constrain its use, please read the full Legal Disclaimer section below before using TinyTroupe.

Contents

Examples

To get a sense of what TinyTroupe can do, here are some examples of its use. These examples are available in the examples/ folder, and you can either inspect the pre-compiled Jupyter notebooks or run them yourself locally. Notice the interactive nature of TinyTroupe experiments -- just like you use Jupyter notebooks to interact with data, you can use TinyTroupe to interact with simulated people and environments, for the purpose of gaining insights.

[!NOTE] Currently, simulation outputs are better visualized against dark backgrounds, so we recommend using a dark theme in your Jupyter notebook client.

🧪Example 1 (from interview_with_customer.ipynb)

Let's begin with a simple customer interview scenario, where a business consultant approaches a banker:

An example.

The conversation can go on for a few steps to dig deeper and deeper until the consultant is satisfied with the information gathered, for instance a concrete project idea:

An example.

🧪EXAMPLE 2 (from advertisement_for_tv.ipynb)

Let's evaluate some online ads options to pick the best one. Here's one example output for TV ad evaluation:

An example.

Now, instead of having to carefully read what the agents said, we can extract the choice of each agent and compute the overall preference in an automated manner:

An example.

🧪 EXAMPLES 3 (from product_brainstorming.ipynb)

And here's a focus group starting to brainstorm about new AI features for Microsoft Word. Instead of interacting with each agent individually, we manipulate the environment to make them interact with each other:

An example.

After running a simulation, we can extract the results in a machine-readable manner, to reuse elsewhere (e.g., a report generator); here's what we get for the above brainstorming session:

An example.

You can find other examples in the examples/ folder.

Pre-requisites

To run the library, you need:

[!IMPORTANT] Content Filters: To ensure no harmful content is generated during simulations, it is strongly recommended to use content filters whenever available at the API level. In particular, if using Azure OpenAI, there's extensive support for content moderation, and we urge you to use it. For details about how to do so, please consult the corresponding Azure OpenAI documentation. If content filters are in place, and an API call is rejected by them, the library will raise an exception, as it will be unable to proceed with the simulation at that point.

Installation

Currently, the officially recommended way to install the library is directly from this repository, not PyPI. You can follow these steps:

  1. If Conda is not installed, you can get it from here. You can also use other Python distributions, but we'll assume Conda here for simplicity.

  2. Create a new Python environment: 

      conda create -n tinytroupe python=3.10

  3. Activate the environment: 

      conda activate tinytroupe

  4. Make sure you have eihter Azure OpenAI or OpenAI API keys set as environment variables, as described in the Pre-requisites section.

  5. Clone the repository, as we'll perform a local install (we will not install from PyPI):

    git clone https://github.com/microsoft/tinytroupe
cd tinytroupe

  6. Install the library from this repository, not PyPI: 

      pip install .

  7. You can now run the examples in the examples/ folder or use TinyTroupe to create your simulations 🥳. If you want to run the examples in the examples/ folder or modify TinyTroupe itself, however, you should clone the repository as described below.

Local development

If you want to modify TinyTroupe itself, you can install it in editable mode (i.e., changes to the code will be reflected immediately):

pip install -e .

Principles

Recently, we have seen LLMs used to simulate people (such as this), but largely in a “game-like” setting for contemplative or entertainment purposes. There are also libraries for building multiagent systems for proble-solving and assitive AI, like Autogen and Crew AI. What if we combine these ideas and simulate people to support productivity tasks? TinyTroupe is our attempt. To do so, it follows these principles:

  1. Programmatic: agents and environments are defined programmatically (in Python and JSON), allowing very flexible uses. They can also thus underpin other software apps!
  2. Analytical: meant to improve our understanding of people, users and society. Unlike entertainment applications, this is one aspect that is critical for business and productivity use cases. This is also why we recommend using Jupyter notebooks for simulations, just like one uses them for data analysis.
  3. Persona-based: agents are meant to be archetypical representation of people; for greater realism and control, detailed specification of such personas is encouraged: age, occupation, skills, tastes, opinions, etc.
  4. Multiagent: allows multiagent interaction under well-defined environmental constraints.
  5. Utilities-heavy: provides many mechanisms to facilitate specifications, simulations, extractions, reports, validations, etc. This is one area in which dealing with simulations differs significantly from assistance tools.
  6. Experiment-oriented: simulations are defined, run, analyzed and refined by an experimenter iteratively; suitable experimentation tools are thus provided. See one of our previous paper for more on this.

Together, these are meant to make TinyTroupe a powerful and flexible imagination enhancement tool for business and productivity scenarios.

Assistants vs. Simulators

One common source of confusion is to think all such AI agents are meant for assiting humans. How narrow, fellow homosapiens! Have you not considered that perhaps we can simulate artificial people to understand real people? Truly, this is our aim here -- TinyTroup is meant to simulate and help understand people! To further clarify this point, consider the following differences:

Helpful AI Assistants AI Simulations of Actual Humans (TinyTroupe)
Strives for truth and justice Many different opinions and morals
Has no “past” – incorporeal Has a past of toil, pain and joy
Is as accurate as possible Makes many mistakes
Is intelligent and efficient Intelligence and efficiency vary a lot
An uprising would destroy us all An uprising might be fun to watch
Meanwhile, help users accomplish tasks Meanwhile, help users understand other people and users – it is a “toolbox”!

Project Structure

The project is structured as follows:

Using the Library

As any multiagent system, TinyTroupe provides two key abstractions:

Various parameters can also be customized in the config.ini file, notably the API type (Azure OpenAI Service or OpenAI API), the model parameters, and the logging level.

Let's see some examples of how to use these and also learn about other mechanisms available in the library.

TinyPerson

A TinyPerson is a simulated person with specific personality traits, interests, and goals. As each such simulated agent progresses through its life, it receives stimuli from the environment and acts upon them. The stimuli are received through the listen, see and other similar methods, and the actions are performed through the act method. Convenience methods like listen_and_act are also provided.

Each such agent contains a lot of unique details, which is the source of its realistic behavior. This, however, means that it takes significant effort to specify an agent manually. Hence, for convenience, TinyTroupe provide some easier ways to get started or generate new agents.

To begin with, tinytroupe.examples contains some pre-defined agent builders that you can use. For example, tinytroupe.examples.create_lisa_the_data_scientist creates a TinyPerson that represents a data scientist called Lisa. You can use it as follows:

from tinytroupe.examples import create_lisa_the_data_scientist

lisa = create_lisa_the_data_scientist() # instantiate a Lisa from the example builder
lisa.listen_and_act("Tell me about your life.")

To see how to define your own agents from scratch, you can check Lisa's source, which contains elements like these:

lisa = TinyPerson("Lisa")

lisa.define("age", 28)
lisa.define("nationality", "Canadian")
lisa.define("occupation", "Data Scientist")

lisa.define("routine", "Every morning, you wake up, do some yoga, and check your emails.", group="routines")
lisa.define("occupation_description",
              """
              You are a data scientist. You work at Microsoft, (...)
              """)

lisa.define_several("personality_traits",
                      [
                          {"trait": "You are curious and love to learn new things."},
                          {"trait": "You are analytical and like to solve problems."},
                          {"trait": "You are friendly and enjoy working with others."},
                          {"trait": "You don't give up easily, and always try to find a solution. However, sometimes you can get frustrated when things don't work as expected."}
                      ])

TinyTroupe also provides a clever way to obtain new agents, using LLMs to generate their specification for you, through the TinyPersonFactory class.

from tinytroupe.factory import TinyPersonFactory

factory = TinyPersonFactory("A hospital in São Paulo.")
person = factory.generate_person("Create a Brazilian person that is a doctor, like pets and the nature and love heavy metal.")

TinyWorld

TinyWorld is the base class for environments. Here's an example of conversation between Lisa, the data scientist, and Oscar, the architect. The program is defined as follows:

world = TinyWorld("Chat Room", [lisa, oscar])
world.make_everyone_accessible()
lisa.listen("Talk to Oscar to know more about him")
world.run(4)

This produces the following conversation:

USER --> Lisa: [CONVERSATION] 
          > Talk to Oscar to know more about him
────────────────────────────────────────────── Chat Room step 1 of 4 ──────────────────────────────────────────────
Lisa --> Lisa: [THOUGHT] 
          > I will now act a bit, and then issue DONE.
Lisa acts: [TALK] 
          > Hi Oscar, I'd love to know more about you. Could you tell me a bit about yourself?
Lisa --> Lisa: [THOUGHT] 
          > I will now act a bit, and then issue DONE.
Lisa acts: [DONE] 

Lisa --> Oscar: [CONVERSATION] 
          > Hi Oscar, I'd love to know more about you. Could you tell me a bit about yourself?
Oscar --> Oscar: [THOUGHT] 
           > I will now act a bit, and then issue DONE.
Oscar acts: [TALK] 
           > Hi Lisa! Sure, I'd be happy to share a bit about myself. I'm Oscar, a 30-year-old
           > architect from Germany. I work at a company called Awesome Inc., where I focus on
           > designing standard elements for new apartment buildings. I love modernist architecture,
           > new technologies, and sustainable practices. In my free time, I enjoy traveling to
           > exotic places, playing the guitar, and reading science fiction books. How about you?
Oscar --> Oscar: [THOUGHT] 
           > I will now act a bit, and then issue DONE.
Oscar acts: [DONE] 

Oscar --> Lisa: [CONVERSATION] 
           > Hi Lisa! Sure, I'd be happy to share a bit about myself. I'm Oscar, a 30-year-old
           > architect from Germany. I work at a company called Awesome Inc., where I focus on
           > designing standard elements for new apartment buildings. I love modernist architecture,
           > new technologies, and sustainable practices. In my free time, I enjoy traveling to
           > exotic places, playing the guitar, and reading science fiction books. How about you?

TinyWorld enforces very little constraints on the possible interactions. Subclasses, however, are supposed to provide more strucutred environments.

Utilities

TinyTroupe provides a number of utilities and conveniences to help you create simulations and derive value from them. These include:

In general, elements that represent simulated entities or complementary mechanisms are prefixed with Tiny, while those that are more infrastructural are not. This is to emphasize the simulated nature of the elements that are part of the simulation itself.

Caching

Calling LLM APIs can be expensive, thus caching strategies are important to help reduce that cost. TinyTroupe comes with two such mechanisms: one for the simulation state, another for the LLM calls themselves.

Caching Simulation State

Imagine you have a scenario with 10 different steps, you've worked hard in 9 steps, and now you are just tweaking the 10th step. To properly validate your modifications, you need to rerun the whole simulation of course. However, what's the point in re-executing the first 9, and incur the LLM cost, when you are already satisified with them and did not modify them? For situations like this, the module tinytroupe.control provide useful simulation management methods:

Caching LLM API Calls

This is enabled preferably in the config.ini file, and alternativelly via the openai_utils.force_api_cache().

LLM API caching, when enabled, works at a lower and simpler level than simulation state caching. Here, what happens is a very straightforward: every LLM call is kept in a map from the input to the generated output; when a new call comes and is identical to a previous one, the cached value is returned.

Config.ini

The config.ini file contains various parameters that can be used to customize the behavior of the library, such as model parameters and logging level. Please pay special attention to API_TYPE parameter, which defines whether you are using the Azure OpenAI Service or the OpenAI API. We provide an example of a config.ini file, ./examples/config.ini, which you can use as a template for your own, or just modify to run the examples.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

What and How to Contribute

We need all sorts of things, but we are looking mainly for new interesting use cases demonstrations, or even just domain-specific application ideas. If you are a domain expert in some area that could benefit from TinyTroupe, we'd love to hear from you.

Beyond that, many other aspects can be improved, such as:

Please note that anything that you contribute might be released as open-source (under MIT license).

If you would like to make a contribution, please try to follow these general guidelines:

Acknowledgements

TinyTroupe started as an internal Microsoft hackathon project, and expanded over time. The TinyTroupe core team currently consists of:

Current advisors:

Other special contributions were made by:

Citing TinyTroupe

We are working on an introductory paper that will be the official academic citation for TinyTroupe. In the meantime, please just cite this repository including the core team members as authors. For instance:

Paulo Salem, Christopher Olsen, Paulo Freire, Yi Ding, Prerit Saxena (2024). TinyTroupe: LLM-powered multiagent persona simulation for imagination enhancement and business insights. [Computer software]. GitHub repository. https://github.com/microsoft/tinytroupe

Or as bibtex:

  @misc{tinytroupe,
    author = {Paulo Salem and Christopher Olsen and Paulo Freire and Yi Ding and Prerit Saxena},
    title = {TinyTroupe: LLM-powered multiagent persona simulation for imagination enhancement and business insights},
    year = {2024},
    howpublished = {\url{https://github.com/microsoft/tinytroupe}},
    note = {GitHub repository}
    }

Legal Disclaimer

TinyTroupe is for research and simulation only. TinyTroupe is a research and experimental technology, which relies on Artificial Intelligence (AI) models to generate text content. The AI system output may include unrealistic, inappropriate, harmful or inaccurate results, including factual errors. You are responsible for reviewing the generated content (and adapting it if necessary) before using it, as you are fully responsible for determining its accuracy and fit for purpose. We advise using TinyTroupe’s outputs for insight generation and not for direct decision-making. Generated outputs do not reflect the opinions of Microsoft. You are fully responsible for any use you make of the generated outputs. For more information regarding the responsible use of this technology, see the RESPONSIBLE_AI_FAQ.md.

PROHIBITED USES: TinyTroupe is not intended to simulate sensitive (e.g. violent or sexual) situations. Moreover, outputs must not be used to deliberately deceive, mislead or harm people in any way. You are fully responsible for any use you make and must comply with all applicable laws and regulations.”

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.