enact

enact - A Framework for Generative Software.

Enact is a python framework for building generative software, specifically software that integrates with machine learning models or APIs that generate distributions of outputs.

Key features

Enact was created to help developers build generative software. Our focus is on offering core framework functionality required for developing, inspecting and improving software that samples from generative subsystems such as LLMs and diffusion models. See why-enact for a more in-depth discussion. To this end, enact boosts your existing python programs by giving you the ability to

• recursively track generative executions,
• persist execution data and system parameters in versioned storage,
• explore the space of possible system executions via rewind and replay,
• turn your existing code into asynchronous flows that can halt and sample input from users, and
• orchestrate higher-order generative flows in which AI systems write and run software.

Enact represents a first-principles approach to supporting development of generative software close to the level of the programming language. It is therefore different from and complementary to libraries that focus on unified integrations (e.g., API wrappers) or on particular algorithm schemas for orchestrating generative flows (e.g., chain/tree-of-thought based approaches).

Installation and overview

Enact is available as a pypi package and can be installed via:

pip install enact

Invocations

Enact recursively tracks inputs and outputs of functions registered with the framework. A record of an execution is called an Invocation, and can be persisted to a store. This is useful for tracking the behavior generative software.

import enact
import random

@enact.register
def roll_die(sides: int) -> int:
  """Roll a die."""
  return random.randint(1, sides)

@enact.register
def roll_sum(num_rolls: int) -> int:
  """Roll dice."""
  return sum(roll_die(6) for _ in range(num_rolls))

with enact.InMemoryStore() as store:
  invocation = enact.invoke(roll_sum, (2,))
  print(enact.invocation_summary(invocation))

Output:

->roll_sum(2) = 7
  ->roll_die(6) = 2
  ->roll_die(6) = 5

Rewind/Replay

Invocations can be rewound and replayed:

# Rewind the last roll and replay it.
with store:
  first_roll = invocation.rewind(1)    # Rewind by one roll.
  print('Partial invocation: ')
  print(enact.invocation_summary(first_roll))
  reroll_second = first_roll.replay()  # Replay the second die roll.
  print('\nReplayed invocation: ')
  print(enact.invocation_summary(reroll_second))

Output:

Partial invocation:
->roll_sum(2) incomplete
  ->roll_die(6) = 2

Replayed invocation:
->roll_sum(2) = 8
  ->roll_die(6) = 2
  ->roll_die(6) = 6

Human-in-the-loop

In enact, humans are treated as just another generative distribution. Their input can be sampled using enact.request_input.

@enact.register
def human_rolls_die():
  """Query the user for a die roll."""
  return enact.request_input(requested_type=int, for_value='Please roll a die.')

@enact.register
def roll_dice_user_flow():
  """Request the number of die to roll and optionally sample die rolls."""
  num_rolls = enact.request_input(requested_type=int, for_value='Total number of rolls?')
  return sum(human_rolls_die() for _ in range(num_rolls))

request_responses = {
  'Total number of rolls?': 3,  # Roll 3 dice.
  'Please roll a die.': 6,      # Humans always roll 6.
}

with store:
  invocation_gen = enact.InvocationGenerator.from_callable(
    roll_dice_user_flow)
  # Process all user input requests in order.
  for input_request in invocation_gen:
    invocation_gen.set_input(request_responses[input_request.for_value()])
  print(enact.invocation_summary(invocation_gen.invocation))

Output:

->roll_dice_user_flow() = 18
  ->RequestInput(requested_type=<class 'int'>, context=None)(Total number of rolls?) = 3
  ->human_rolls_die() = 6
    ->RequestInput(requested_type=<class 'int'>, context=None)(Please roll a die.) = 6
  ->human_rolls_die() = 6
    ->RequestInput(requested_type=<class 'int'>, context=None)(Please roll a die.) = 6
  ->human_rolls_die() = 6
    ->RequestInput(requested_type=<class 'int'>, context=None)(Please roll a die.) = 6

Making your types enact compatible

Enact has built-in support for basic python datatypes such as primitives, lists, sets, dictionaries and tuples.

If you need support for additional types, you can either define them as Resource dataclasses in the framework or define a wrapper.

Defining a resource

import dataclasses

@enact.register
@dataclasses.dataclass
class MyResource(enact.Resource):
  x: int
  y: list = dataclasses.field(default_factory=list)

with store:
  # Commit your resource to the store, obtain a reference.
  ref = enact.commit(MyResource(x=1, y=[2, 3]))
  # Check out your reference.
  print(ref.checkout())  # Equivalent to "print(ref())".

Output:

MyResource(x=1, y=[2, 3])

Defining a wrapper

For existing types, it can be more convenient to wrap them rather than to redefine them as enact Resource objects.

This can be accomplished by registering a TypeWrapper subclass.

class Die:
  """Non-enact python type."""
  def __init__(self, sides):
    self.sides = sides

  @enact.register
  def roll(self):
    return random.randint(1, self.sides)

@enact.register
@dataclasses.dataclass
class DieWrapper(enact.TypeWrapper[Die]):
  """Wrapper for Die."""
  sides: int

  @classmethod
  def wrapped_type(cls):
    return Die

  @classmethod
  def wrap(cls, value: Die):
    """Translate to enact wrapper."""
    return DieWrapper(value.sides)

  def unwrap(self) -> Die:
    """Translate to basic python type."""
    return Die(self.sides)

with store:
  die = Die(sides=6)
  invocation = enact.invoke(die.roll)
  print(enact.invocation_summary(invocation))

Output:

-><__main__.Die object at 0x7f3464398340>.roll() = 4

Documentation

Full documentation is work in progress. See enact concepts for more information and take a look at the examples.

Why enact - A manifesto

Generative AI models are transforming the software development process.

Traditional software relies primarily on functional buildings blocks in which inputs and system state directly determine outputs. In contrast, software that samples one or more generative subsystems associates each input with a range of possible outputs.

This small change in emphasis - from functions to conditional distributions - implies a shift across multiple dimensions of the engineering process, summarized in the table below.

	Generative	Traditional
Building block	Conditional distributions	Deterministic functions
Engineering	Improve output distributions	Add features, debug errors
Subsystems	Unique	Interchangeable
Code sharing	Components	Frameworks
Executions	Training data	Logged for metrics/debugging
Interactivity	Within system components	At system boundaries
Code vs data	Overlapping	Distinct

Building generative software means fitting distributions

Generative AI can be used to quickly sketch out impressive end-to-end experiences. Prototypes of text-based agents, for instance, often involve little more than calling an API with a well-chosen prompt. Evolving such prototypes into production-ready systems can be non-trivial though, as they may show undesirable behaviors a certain percentage of the time, may fail to respond correctly to unusual user inputs or may simply fail to meet a softly defined quality target.

Where traditional engineering focuses on implementing features and fixing bugs, generative software engineering tunes the distribution represented by the system as a whole towards some implicitly specified target, e.g., a set of example outputs, corrections of previous outputs or task rewards.

In some cases, tuning generative systems can be achieved directly using machine learning techniques and frameworks. For instance, systems that consist of a thin software wrapper around a monolithic machine learning model may directly train the underlying model to shape system behavior.

In other cases, particularly when running complex generative flows that involve multiple interacting subcomponents that cannot be easily optimized end-to-end, gradient-free methods are necessary to improve system performance. These range from using programmers (either human or synthetic) to elaborate the algorithmic guard-rails of the generative process, running experiments between API identical subsystems or finding ideal parameters (e.g., prompt templates) for individual components.

Recursive elaboration and end-to-end compression

In generative software, individual components produce distributions that may be more or less well-suited to the system's overall goal: An LLM that performs well on tabular data may perform less well when used as a chat agent. One fine-tuned diffusion model may produce images that are closer to the target style than another. This is distinct from the the case of traditional software engineering, where the choice between two correct, API-identical implementations of a module is of little importance to the behavior of the system as a whole.

In addition to a choice between different machine learning models, there is a choice between multiple algorithmic elaborations of a given model. The behavior of GPT, for example, can be improved by providing algorithmic guardrails that structure and steer the generation process such as chain-of-thought or tree-of-thought prompting.

The profusion of base models and possible algorithmic elaborations of base models suggests that gradient-free search methods will take a central place in generative software development. This could be as simple as running an experiment in which one component is compared against another or as complex as searching the space of software, e.g., using LLM-boosted genetic algorithms.

Once a component has been optimized using such methods, the resulting data can be used in gradient-based optimization to 'compress' the gained knowledge into end-to-end trained models. This suggests that generative systems will alternate gradient-based and gradient-free methods, with the former supplying new training signal for the latter. (This loop is the core insight behind algorithms in the AlphaZero family, in which data from a search-based elaboration of a model is used to train the next iteration of the model.)

Generative software is collaborative

There are many generative AI components that are mutually API-compatible (e.g., text-to-image generation, LLMs) but that cannot be directly ranked in terms of quality since they represent a unique take on the problem domain. The combination of API compatibility and variation lends itself to a more evolutionary, collaborative style than traditional software, where shared effort tends to centralize into fewer, lower-level frameworks.

This new collaborative approach is already evidenced by both widespread sharing of prompt templates and the existence of public databases of fine-tuned models.

Generative software reflects on its executions.

Many popular programming languages offer capacity for reflection, wherein the structure of code is programmatically interpretable by code. Generative software additionally requires deep reflection, that is, the ability for code to reflect on code execution.

Sampling a generative model is computationally intensive and provides valuable information that may be used as system feedback or training data. Since generative software may include complex, recursive algorithms over generative components, this suggests a need for not simply tracking inputs and outputs at system boundaries, but at every level of execution.

This is particularly relevant to higher order generative flows, in which the output of one generative step structures the execution of the next: Such a system can use call traces as feedback to improve its output, in the same way that a human may use a debugger to step through an execution to debug an issue.

Humans are in the loop

Unlike traditional systems, where user interactions happen at system boundaries, AI and humans are often equal participants in generative computations. An example is Reinforcement Learning from Human Feedback (RLHF), where humans provide feedback on AI output, only to be replaced by an AI model that mimics their feedback behavior during the training process. In other cases, critical generative flows (e.g., the drafting of legal documents) may require a human verification step, without which the system provides subpar results.

The choice between sampling human input and sampling a generative model is involves considerations such as cost and quality, and may be subject to change during the development of a system. For example, an early deployment of a system may heavily sample outputs or feedback from human participants, until there is enough data to bootstrap the target distribution using automated generative components.

Therefore, a generative system should be designed to utilize human input in the same way it would interact with an API or subsystem.

Data is code, code is data

Traditional software systems tend to allow a clear distinction between code (written by developers) and data (generated by the user and the system). In generative systems this distinction breaks down: Approaches such as AutoGPT or Voyager use generative AI to generate programs (specified in code or plain text), which in turn may be interpreted by generative AI systems; prompts for chat-based generative AI could equally be considered code or data.

Development and Contributing

The framework is open source and Apache licensed. We are actively looking for contributors that are excited about the vision. If you're interested in the project, feel free to reach out to [e-mail].

You can download the source code, report issues and create pull requests at https://github.com/agentic-ai/enact.