Credit Card Statement Categorizer using AI (private, no cloud)

A local (no internet required) Large Language Model (LLM) performs the task of reading credit card transaction descriptions (messy text) and assigning the best category out of a list you give it.

Why

Showcase an advanced NLP use case of LLMs with no subscriptions or privacy concerns like one might have when through a third party cloud api.
To show an example of how to setup a local LLM comparable to GPT3.
Demonstrate prompt engineering an LLM (and sanitizing the output) to do your bidding. Uses a GUI to create a useful prompt.
Give an example of rapid prototyping something to get an idea if it is worth polishing.

Optionally enter some examples via the GUI window and click the button. A custom prompt will be created and sent to the LLM.
Timing data and debug info is printed to the console as the spreadsheet file is processed line by line.
It outputs a new results.csv spreadsheet with a column "Category_LLM" containing the category selected by the LLM.

This is a prototype to display the technique, nothing more and not production code.

Download the mistral-7b or similar LLM model file (Google: "mistral-7b-instruct-v0.2.Q4_K_M.gguf huggingface") from huggingface and put it in its own directory.
Change the folder /home/bb/llms in main.py to point to your model file. (mistral-7b recommended, but others work!)
python3.10 -m venv myenv # create virtual environment
source myenv/bin/activate # activate virtual environment
Install python package llama-cpp-python with or without cuda support
To install with cuda:
- install cmake if needed, e.g. apt-get install cmake on linux
- FORCE_CMAKE=1 CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python
If you want to use CPU only, (10-30x slower per answer, but still delivers the point) just skip the line above
Install the other dependencies:
pip install -r requirements.txt
Run with python main.py

If you found this helpful and got to understand LLMs prototyping a bit better, I would be extremely glad if you could give it a star! ⭐