epec254 / gpt-intuition

gpt-intuition is a streamlit-based embedding visualization tool designed to help ML practitioners gain intuition into how their data maps into embedding spaces. Inspired by Google's tensorflow projector.

Embedding Projector for Streamlit

Context

Embeddings are the core of modern machine learning (including LLMs and other large transformer based models). I believe a critical step in using LLMs (or any other model for that instance) is developing intuition around how the embedding spaces represent your data.

While OpenAI provides a Notebook for visualizing embeddings, I believe more robust tools are needed. Tensorflow's Embedding Projector is an underrated tool that fills this gap, but can be painful to use (need to export embeddings into a specific set of 2 files, must be loaded as a standalone application, etc).

This is a "good enough" version of a visualization tool based on Streamlit and open source uMAP and tSNE algorithms. The tool's primary goal is to make it easy for you to load your own embeddings quickly.

Working demo version

Courtesy of Streamlit cloud, here is a working demo: https://gpt-intuition-embedding-visualizer.streamlit.app/

Usage

I use pipenv for management of pip packages & virtualenvs. You can either do pipenv install or pip install -r requirements.txt. If you use pip directly, I suggest doing so inside a virtual environment (pipenv handles this for you).

Run streamlit run embed.py to launch the Streamlit web application.

Configuration

The app comes configured with 4 sample sets of embeddings, 2 shamelessly stolen from Tensorflow and 2 generated with llama-index and langchain from the Paul Graham essay.

Adding your own is easy:

1. Create a CSV file in the /data/ directory with the following columns:
   1. Required a set of columns "embedding_1", "embedding_2", ... that represent the embedding itself
   2. Optional label or other metadata column (e.g., name that each the embedding represents, etc)
   3. Optional label to use for coloring the points on the graph (e.g., a low dimensional bucket-esque label)
2. Add a dict in the same format as the one below to DATA_SETS in embed.py. Dimensions is required to properly load the embeddings.

DATA_SETS = {
    ...
    "word2vec-10k-sample": {
        "data_file": "datasets/word2vec_10000_200d_merged.csv", # required
        "dimensions": 200, # required
        "label_column": "word", # optional
        "color_column": "", # optional
    },
}

Roadmap

• Upload CSV file and visualize
• Add a way to select a single embedding and see the closet KNN within the space
• Add in support for generating OpenAI-based embeddings directly from the UX
• Automatically load any CSV file put into /data/
• Add support for creating a GPT-Index or LangChain's chains and visually seeing how it works
• Animations!!
• Testing - aka anything more than refresh and check ;-)

Known issues

• Sometimes, the app exits with the error Terminating: Nested parallel kernel launch detected, the workqueue threading layer does not supported nested parallelism. Try the TBB threading layer. - I believe this is due to the uMAP algorithm's implemention of multi-threading conflicting with Streamlit, but have not debugged the issue in depth. Typically, restarting it a few times works.

Credits

Code for the uMAP and tSNE implementations with Plotly courtesy of the Plotly docs.