openbrain 0.20.0

A package to interact with generative AI and build specialized generative AI workflows. This project is dual-licensed under AGPL-3.0 and a separate commercial license.

OpenBrain

🚧 Under active development 🚧

OpenBrain is a tool-wielding, cloud native, LLM agent platform. It provides APIs and tools to configure, store, and retrieve LangChain agents, making your chat sessions and workflows stateful and persistent. You will find in-memory and DynamoDB ORM options as well as a SAM template for deploying the necessary resources for stateless agents to your AWS. The use of these mixins is controlled by environment variables.

OpenBrain agents are stateful by nature, so they can remember things about you and your conversation. They can also use tools, so you can use the same agent to chat and to perform actions. This project provides a mechanisms to integrate with an API to store the state of the agent as a session, so that the agent can be used asynchronously from any source in a serverless environment.

Interactions with the agent can be injected into any application easily by constructing a query, choosing an agent configuration, and pre-processing your data through that agent before sending it of for further processing.

Openbrain as a Service

To test Openbrain as a service, please do the following:

1. Register at https://portal.openbra.in.
2. Log in to the portal and subscribe to the Openbrain service.
3. Navigate to your dashboard in the portal and use your API keys to interact with the API.
4. A swagger UI and SDKs are available for subscribers.

NOTE: There is currently no fee for using the service, but it's using my personal AWS and OpenAI accounts, so I'll pull the plug immediately if it becomes expensive.

Features

• Interactive Agent Tuner: A GUI to modify and test agent configurations.
• Command-Line Interface: Use ob for quick completions and ob-chat for an interactive session.
• Flexible Configuration: Customizable agents through DynamoDB backed ORM.
• Event-Driven Architecture: Extensible through cloud-based event handling.

Quick Start

# TLDR: demo functionality
pip install openbrain[gradio]
ob-tuner

To print an example .env file, use ob-env.

ob-env

Installation

Gradio is included as an extra dependency. This keeps the base installation small enough to fit in a lambda layer. If you intend to experiment with the ob-tuner interface locally, add the [gradio] extra to your pip install command.

pip install openbrain[gradio]

Setup .env file

cp .env.example .env  # Edit this file with your own values

Deploy Supporting Infrastructure

The AgentConfigs you create in ob-tuner can be saved to your DynamoDB tables in AWS, allowing you to reference them in any AWS workflow using the boto3 library.

If you intend to use this project in your API, an AWS SAM template is provided to get you started. You can deploy the template using the following command. You will need to have the AWS CLI installed and configured with your credentials.

:warning: This will deploy resources to your AWS account. :warning: You will be charged for these resources. See AWS Pricing for more information.

python ci_cd.py -I

Using OpenBrain

OpenBrain Gradio Tuner

To facilitate tuning agent parameters and experimenting with prompts, OpenBrain provides a GUI interface using Gradio. You can use the in-memory object model to store your agent configurations locally. This is controllable by setting GRADIO_LOCAL=True, ORM_LOCAL=True, UTIL_LOCAL=True in your .env file.

$ ob-tuner
Running on local URL:  http://0.0.0.0:7861

To create a public link, set `share=True` in `launch()`.
Tip: You can show or hide the button for flagging with the `allow_flagging=` kwarg; for example: gr.Interface(..., allow_flagging=False)

Command Line Completions

(doesn't work quite like this yet, but that's the idea, currently noisy)

$ ob --profile_name=linux 'change the file permissions of all files that start with a "w" in this directory (recursively) to 042'

find . -type f -name 'w*' -exec chmod 042 {} +

Command Line Interactive Session

(doesn't work quite like this yet, but that's the idea)

$ ob-chat
---------- Begin Chat Session ----------
> OpenBrain: What… is your name?
User: It is Arthur – King of the Britons.
> OpenBrain: What… is your quest?
User: To seek the Holy Grail.
> OpenBrain: What… is the air-speed velocity of an unladen swallow?
User: What do you mean? An African or a European swallow?
> OpenBrain: I don’t know that. Aaaaaaaaagh!

Architecture Overview

Data Model

classDiagram
    class ChatMessage {
        User provided*
        + str: sessionId
        + str: clientId
        + str: message
        + AgentConfig: agentConfigOverrides
        + str: agentConfig
        + bool: reset
    }

    class AgentConfig {
        + str: profileName
        + str: clientId
        + str: iceBreaker
        + Decimal: temperature
        + ...
        + save()
        + refresh()
        + load()
        + delete()
        + get_all()
    }

    class ChatSession {
        + str: sessionId
        + str: clientId
        + byte: agentMemory
        + AgentConfig: agentConfig
        + Lead: lead
        + save()
        + load()
    }

    class Lead {
        + str: sessionId
        + str: clientId
        + str: fullName
        + str: firstName
        + str: middleName
        + str: lastName
        + List[str]: medications
        + str: email
        + ...
        + save()
        + refresh()
        + delete()
        + load()
        + get_all()
    }

    ChatSession "1" *-- "1" Lead: contains
    ChatSession "1" *-- "1" AgentConfig: contains
%%    ChatMessage "1" *-- "1" AgentConfig: contains
%%    ChatSession "1" *-- "*" ChatMessage: contains
    ChatSession "1" *-- "1" langchain_ChatMemory: from langchain, serialized

Data Flow diagram

OpenBrain uses an event driven architecture. The agent sends events to event bus and then the developer can simply write rules and targets for the incoming events once the targets are ready. The following diagram shows the data flow in two parts.

1. The user interaction with the agent and the agent interaction with an event bus.
2. The event bus and the targets that are triggered by the events.

sequenceDiagram
    title Agent Data Flow
    participant User
    create participant GPT Agent
    participant AgentConfigTable
    participant OpenAI
    participant Tool
    participant EventBus

    User ->> GPT Agent: (AgentConfig, AgentMemory), ChatMessage
        GPT Agent -->> AgentConfigTable: profileName
        AgentConfigTable -->> GPT Agent: AgentConfig
        GPT Agent -->> OpenAI: ChatMessage
        OpenAI -->> GPT Agent: ChatMessage
        GPT Agent -->> Tool: Tool(Object, clientId)
        Tool -->> EventBus: (Object, clientId, session_id, object_id)
        Tool -->> GPT Agent: ChatMessage
    destroy GPT Agent
    GPT Agent ->> User: ChatMessage, (AgentConfig, AgentMemory), Object

  box blue Databases
      participant AgentConfigTable
  end
  box purple Tool
      participant Tool
  end

  box gray EventBus
      participant EventBus
  end

  box red Provider
      participant OpenAI
  end

sequenceDiagram
    title Agent Data Flow
    participant SQS
    participant EventBus
    participant Lambda
    participant ObjectTable
    participant AgentConfigTable
    participant ChatHistoryTable
    participant ExternalSite

    EventBus ->> Lambda: (Object, clientId, sessionId, objectId)
    Lambda -->> ObjectTable: (clientId, objectId)
    ObjectTable -->> Lambda: Object

    Lambda -->> AgentConfigTable: (profileName, clientId)
    ChatHistoryTable -->> Lambda: AgentConfig

    Lambda --> ChatHistoryTable: (clientId, sessionId)
    ChatHistoryTable -->> Lambda: (AgentMemory, AgentConfig)

    Lambda ->> ExternalSite: ...
    ExternalSite --x Lambda: ERROR
    Lambda ->> SQS: <DETAILS NEEDED TO RETRY>
    ExternalSite ->> Lambda: ...

    Lambda -> EventBus: <POTENTIAL NEW EVENT>

    box maroon DeadLetterQueue
        participant SQS
    end

    box blue Databases
        participant ObjectTable
        participant AgentConfigTable
        participant ChatHistoryTable
    end

    box gray EventBus
        participant EventBus
    end

    box brown EventTargets
        participant Lambda
    end

    box green Internet
        participant ExternalSite
    end

Contributing

See CONTRIBUTING.md for guidelines.

License

• Open Source: AGPLv3
• Commercial: See COMMERCIAL_LICENSE and contact us for inquiries.