auto-graph-of-thoughts 1.1.0

Automated Graph of Thoughts

Automated Graph of Thoughts

This is the official implementation of Automated Graph of Thoughts.

Setup Guide

To run this code, Python 3.11 or newer is required. The latest version of the package can be installed from PyPI:

pip install auto-graph-of-thoughts

Alternatively, the package can be installed from source.

Optional Dependencies

The project comes with optional dependencies which are required for some features.

Graph Visualization

To visualize the graphs by using pure_graph_of_thoughts.visualization, the optional visualization dependencies are required.

pip install auto-graph-of-thoughts[visualization]

For a cleaner, hierarchical visualization, add dot-visualization.

pip install auto-graph-of-thoughts[visualization,dot-visualization]

Be aware that dot-visualization requires the GraphViz library to be installed.

Notebooks

Several notebooks with examples and model training are provided with the source code. To run the notebooks, the optional notebooks dependencies are required.

pip install auto-graph-of-thoughts[notebooks]

Be aware that the notebooks are not part of the distributed package on PyPI.

Pure Graph of Thoughts

The package pure_graph_of_thoughts contains a new implementation of the Graph of Thoughts concepts.

Graph of Thoughts was originally introduced in the paper Graph of Thoughts: Solving Elaborate Problems with Large Language Models. The official implementation of the paper's proposed API can be found here: https://github.com/spcl/graph-of-thoughts.

The pure_graph_of_thoughts package does not conform the API proposed by the original paper nor is it a fork of it. It aims for a more automation-friendly implementation of the general concept of Graph of Thoughts, where both construction and traversal of a graph can be handled iteratively.

Some key differences and restrictions:

• Operations and thoughts are represented independently of their graph structure.
• As a user-facing API, operations can be defined in a declarative way over a typed and validated data structure (DSL).
• There is a strict distinction between a prompt operation executed by a language model and a code execution operation.
• To simplify parsing logic and to ensure consistent results, the JSON format is used for communication with the language model.
• The scoring is now part of an operation involving a prompt, rather than being a standalone operation that can be added arbitrarily. While this simplifies the automation process, it restricts the user's possibility of adding a validation operation.