xMind -- AgentFlow Framework

xMind is a modular framework built with XLang, designed to implement Large Language Model (LLM) Memory, Planning, and Agent-flow capabilities. This project allows developers to seamlessly integrate advanced AI features like context retention, decision-making, and dynamic dataflows into their applications.

Features

• LLM Memory: Retain and utilize context across sessions to enhance interaction and decision-making.
• Planning: Implement sophisticated planning mechanisms that allow LLMs to make informed decisions based on historical data and projected outcomes.
• Agent-flow Management: Orchestrate complex Agent-flows to streamline processing and enhance the performance of AI-driven applications.
• Modular Design: Easily extend and customize the framework to fit your specific needs.

AgentFlow Graph

Getting Started

Prerequisites

• XLang Please clone XLang into the xMind/ThirdParty folder and ensure the folder is named xlang.

Build the Framework

Clone the repository:

git clone https://github.com/xlang-foundation/xMind.git
cd xMind
mkdir build
cd build
cmake ..
make

for xcode, use cmake -G Xcode .. to generate Xcode project

Terms and Concepts

1. Blueprint: A YAML-based structure used to define various elements such as variables, prompts, actions, and more.

2. Variable:

   • Scope: Variables are global within the same file and do not require a prefix.
   • Cross-File Access: When accessing a variable from another file, a prefix must be used, e.g., file1.var1.

3. Node:

   • Represents a component in AgentFlow, using a graph-based approach to connect various nodes.

4. Function:

   • A node within AgentFlow that serves as an inline translate node. It supports only one input and one output.

5. Action:

   • A buffered node in AgentFlow that processes input through a separate thread (in XLang) or a process (in Python).
   • Use Case: Actions are typically used to connect to external environments such as REST APIs, file access, or UIs.

6. Agent:

   • A specialized node within AgentFlow that performs LLM (Large Language Model) inference.
   • Core Node: It serves as the core of AgentFlow, buffering inputs and combining them with prompts from various sources before making an inference request to an LLM.

7. LlmPool:

   • Managed by xMind, this concept involves handling LLM requests in a pool, based on factors like HTTP request status and LLM key usage time limits.

8. Session Memory:

   • Session Persistence: Each chat completion is maintained within a session, ensuring continuity across interactions.
   • Session Identifiers: Externally, each session is identified by a globally unique identifier (GUID). Internally, sessions are tracked using an integer that loops for efficient resource management.
   • Node Data Handling: The first item in each node’s input and output data is the internal session ID. This approach allows a single graph instance to serve multiple chat instances, optimizing resource usage.
   • Session Memory: Sessions maintain a history of interactions as session memory. When making requests, this history is automatically bound as part of the prompt, ensuring context is preserved.
   • LLM Output Integration: All outputs from the language model (LLM) are fed back into the session memory, continuously enriching the session’s context.

Running the Framework

Start Guide

CLI - xmcli

CLI