ds-article-rag

Dataset used

• 1300+ Towards DataScience Medium Articles Dataset (https://www.kaggle.com/datasets/meruvulikith/1300-towards-datascience-medium-articles-dataset)

How to run

Note: Startup of the app as well as LLM querying can take a long time, especially without a GPU.
If you don't have llama2 downloaded in ollama (or your running with docker), first invocation of LLM querying can be very time-consuming due to ollama downloading the llama2 model (it is known to have connection issues).

[16GB RAM, Intel Core i7-9750H CPU, Nvidia GeForce 1660-Ti] -- 3 minutes startup with GPU (no cache) -- 10 minutes startup CPU-only (no cache)

Running with Docker

1. Run git clone https://github.com/AmevinLS/ds-article-rag
2. Change working directory to the cloned repository (cd ds-article-rag)
3. Download dataset and extract it to ./data/medium.csv (create the ./data directory if needed)
4. Run:
   • docker-compose -f docker-compose_cpu.yml up --build to run on CPU, if you don't have an Nvidia GPU
   • docker-compose -f docker-compose_gpu.yml up --build to run using your Nvidia GPU
5. Go to http://localhost:8501 in your browser to open the streamlit app

Running locally

Make sure you have Ollama installed (you can download it here here).
Optionally, you can run ollama pull llama2 yourself to avoid download issues at runtime

1. Run git clone https://github.com/AmevinLS/ds-article-rag
2. Change working directory to the cloned repository (cd ds-article-rag)
3. Download dataset and extract it to ./data/medium.csv (create the ./data directory if needed)
4. Run pip install -r requirements.txt (tested for Python=3.9)
5. Run streamlit run ds_article_rag/app.py
6. Go to http://localhost:8501 in your browser to open the streamlit app

[Optional] Download cache for faster start-up

1. Download 'cache' folder from Google Drive (here)
2. Export it into the ./data, so the resulting structure looks as follows:

data/
  medium.csv
  cache/
    code_reduced_par_embeds.npy
    code_reduced_paragraphs.csv

Demo

Details on the system

You can find the details pertaining to how the whole system is structured and other relevant information in the ./docs/report.md file in this repository