Reforma Pensional QA using Llama 3 (8B-Instruct) and RAG (HF | Langchain+FAISS+HF)

🪒 Setup

Using conda or virtualenv install the packages.

conda create --name <YOUR_ENV_NAME> python=3.10
conda activate <YOUR_ENV_NAME>
pip install -r requirements.txt

Make sure you have at least 12 GB of VRAM.

🐍 Usage

Data preprocessing

Textract was used to get text chunks based on a particular string. These are also saved to a HuggingFace dataset. See link

For more details see src/preprocessing.py

Inference

• HuggingFace See src/inference_hf.py

• Langchain See src/inference_langchain.py

Demo

A Gradio app can be started using python src/app.py

🤿 Contributing to this repo

• This repo uses pre-commit hooks for code formatting and structure. Before adding|commiting changes, make sure you've installed the pre-commit hook running pre-commit install in a terminal. After that changes must be submitted as usual (git add <FILE_CHANGED> -> git commit -m "" -> git push )

• For dependencies, pip-tools is used. Add the latest dependency to the requirements.txt and run pip-compile requirements.txt -o requirements.txt to make sure the requirements file is updated so we can re-install with no package version issues.