Genieversum

Our Mission

Genie Flow (or simply Genie) is a modular AI platform designed to help you build intelligent, reusable agents that can automate tasks, process data, and interact with users or systems. Whether you're a seasoned developer or just starting out, Genie Flow offers a flexible and approachable way to create smart workflows.

Our vision is that a good agentic framework should be:

• Code First – We do not have a "low code" environment. We believe in the power of expression of well-written code.
• Deployable – A Genie agent is ready to deploy as an API. There is no magic engine or deployment formula. Your agent is built into a Docker container and comes with a well-documented API out of the box.
• Scalable – Work is queued to be picked up by workers. This makes your agent immediately scalable. From your laptop to large server clusters. One code base. Scalable out of the box.
• Community Driven – AI should be available for everyone, and everyone should be able to contribute.

We do consulting. We are not a software house. We are not a product company. We are not a service or hosting provider. We make our living by helping large organisations to make the best use of AI. We are the company where Next is Made Real.

What is Genie

Simply put: Genie is an agentic AI orchestrator. It manages the dialogue between an actor (human or machine) and a cascade of calls to external systems. Many of these external systems may be Large Language Models, but the dialogue flow typically combines that with other sources of information and operational systems that can be reached via an API.

Core Concepts

Memory

When you start a new Genie session, it creates a fresh Model instance to hold your agent's properties. This Model is a Pydantic object that you define, and its properties power both external system calls and the responses your users see.

As the conversation unfolds, you decide when and how to update these property values. Behind the scenes, Genie keeps the Model in Redis memory for fast access, then archives it to MongoDB for long-term storage.

Invokers and Renderers

Think of Invokers as your bridge to the outside world. Whether you're calling an LLM, querying a database, processing images, or hitting an API, Invokers handle the heavy lifting. They use Jinja templates to shape your inputs and deliver clean outputs back to your user.

Renderers work the other way around—they take your agent's data and format it into responses for your users. They also use Jinja templates to craft the final output.

Templates

Jinja templates are the backbone of data flow in Genie. They're incredibly flexible, letting you transform data for Invokers and polish responses for users.

For simple cases, just reference a template by name. But when you need more sophisticated data flows, you can chain templates together. Use template lists when you want each template to build on the previous one's output. Use template dictionaries when you need multiple parallel results. And feel free to nest them—mix lists and dictionaries to create powerful multi-stage processing pipelines.

Worker Queue

Genie uses a job queue to keep things running smoothly. Your agent actually runs as two separate processes working in harmony. The API process manages the conversation flow and runs your business logic, while the Worker process handles all the background invocations by pulling jobs from the queue.

Example Genie Agent

Imagine the following use case:

As a user, I want to have a dialogue with a Large Language Model.

Three steps:

1. Create your data model
2. Define how you want the dialogue to flow
3. Specify the templates to use in each of the steps

Defining your data model

First, you define your GenieModel - a Pydantic data model of information that you want to carry during the dialogue session.

class MyFirstModel(GenieModel):

    # there are no specific data elements I need to carry

    # link this model to the state machine
    @classmethod
    def get_state_machine_class(cls) -> type[GenieStateMachine]:
        return GenieStateMachine

Define the flow of your dialogue

Next, you define how your dialogue needs to flow by creating your dialogue's state machine:

class MyFirstMachine(GenieStateMachine):

    # STATES
    into = State(initial=True, value=100)
    ai_creates_response = State(value=200)
    user_enters_query = State(value=300)

    # EVENTS & TRANSITIONS
    user_input = (
        intro.to(ai_creates_response)
        | user_enters_query.to(ai_creates_response)
    )
    ai_extraction = (
        ai_creates_response.to(user_enters_query)
    )

    # TEMPLATES
    templates = dict(
        intro="response/intro.jinja2",
        ai_creates_response="llm/ai_creates_response.jinja2",
        user_enters_query="response/user_enters_query.jinja2",
    )

This creates a dialogue that looks like:

stateDiagram-v2
    direction LR
    [*] --> intro
    intro --> ai_creates_response: user_input
    ai_creates_response --> user_enters_query: processing_done
    user_enters_query --> ai_creates_response: user_input

    intro: Intro
    ai_creates_response: AI Creates Response
    user_enters_query: User Enters Query

The most basic dialogue flow:

1. The Agent introduces themselves and asks a question
2. The user sends their input
3. An LLM formulates a response
4. The engine signals that processing is done
5. The user views the response and sends new input
6. Back to point 3

Creating the templates

We just need to define the templates. First template, the intro.jinja2 template:

Welcome to this simple Question and Answer dialogue Genie Flow example!

How can I help? Please go ahead and ask me anything.

We also need to define the prompt that will get sent to the LLM. This happens in the file ai_creates_response.jinja2:

- role: system
  content: |
    You are a friendly chatbot, aiming to have a dialogue with a human user.
    Your aim is to respond logically, taking the dialogue you had into account.
    Be succinct, to the point, but friendly.
    Stick to the language that the user start their conversation in.

{{ chat_history }}

- role: user
  content: |
{{ actor_input|indent(width=4, first=True) }}

This template defines the system prompt, followed by the {chat_history}, followed by the input from the previous actor. That will be the human user in our case.

And then, the final template, user_enters_query.jinja2:

{{ actor_input }}

Switch on

That's it! This defines a fully fledged AI agent.

Deploying this would mean: create a main.py that tells Genie where your and data model class live, run the worker, run the API and spin up a front-end that talks to your API. We have provided a simple command-line interface that we use in anger during development, and a simple React chat interface that you can get started with.

Where from here?

• Getting Started for a step-by-step guide to run the above example
• Under the Hood for a detailed description of the underlying architecture
• Deployment which describes how to deploy your Genie
• API for the code-generated API of some of the core components

Repositories