LLM Hierarchical Design

LLM Hierarchical Design uses LLMs to facilitate tree-like conversations for top-down hierarchical project structuring.

This tool allows users to branch conversations out into multiple threads while maintaining the context window, aiding in the iterative development of project components.

Motivation

Hierarchical Project Planning

Many systems exhibit a hierarchical structure designed in layers. For example, a programing project could have several modules, with each module containing several functions. This makes iterative hierarchical project planning more natural than linear project planning.

graph TD;
    R[Project Root]-->A[Module A];
    R-->B[Module B];

    A-->A1[Function A1];
    A-->A2[Function A2];
    B-->B1[Function B1];

Loading

Recency Bias

LLMs exhibit a recency bias within their context windows. They tend to prioritize information presented at the end of the input sequence over earlier content. When a hierarchical conversation structure is flattened to a linear structure in a preorder traversal, the root prompt at the beginning of the input has lower information priority, even though it is close to the beginning prompt in the tree.

graph LR;
    R[Project Root]-->A[Module A];
    A-->A1[Function A1];
    A1-->A2[Function A2];
    A2-->B[Module B];
    B-->B1[Function B1];

Loading

Proposed Solution

To solve the issue of linearizing a complicated component tree, this project uses the following structure for the context window when, for example, working on functions in module B.

graph LR;
    S[Project Component Summary]-->R[Project Root]
    R-->B[Module B];
    B-->B1[Function B1];

Loading

Features

• Supports branching conversations for hierarchical project planning.
• Keeps information about the entire conversation tree within the context window.

Prerequisites

Before using this project, ensure you have:

• The uv package manger.
• An openai api key.

Setup

1. Clone the Repository

git clone https://github.com/KennethJAllen/conversation-tree
cd conversation-tree

2. Install Dependencies

Use UV to create the virtual environment.

uv sync

3. Create a .env File

The project requires an OpenAI API key. Create a .env file in the root directory of the project with contents:

OPENAI_API_KEY=your_openai_api_key

if using an Anthropic API key, instead create a .env file in the root directory of the project with contents:

OPENAI_API_KEY=your_anthropic_api_key

Replace your_openai_api_key with your actual OpenAI API key, or your_anthropic_api_key with your actual Anthropic API key.

If both API keys are included, defaults to the openai api key.

How to Run

This project uses streamlit for the UI. To open the UI, use:

streamlit run main.py

Alternatives

OpenAI's ChatGPT and Anthropic's Claude both support an edit feature that can mimic a tree-like conversation by strategically editing queries. However, this only keeps your current branch in the context window. This means that ChatGPT and Claude would be unaware of the details of different components in the project.