Basic local RAG from scratch

Created throughout Day 80-83 of my self-studying journey.

Repo structure

📁pdf-rag-from-scratch
└── 📁dev
    └── dev_preprocess_pdf.ipynb -> preprocess pdf using llama-index
    └── dev_rag.ipynb -> runs rag using llama-index preprocessed pdf
└── lbg_relationship_tnc.pdf
└── lbg_relationship_tnc_locked.pdf
└── preprocess_pdf.py -> preprocess pdf using langchain
└── rag.py -> run local rag chat using langchain preprocessed pdf
└── requirements.txt

Setup

1. git clone https://github.com/divakaivan/pdf-rag-from-scratch.git
2. pip install -r requirements.txt
3. python preprocess_pdf.py -> PDF must be saved in the same directory as the file, then it reads and processes the PDF for you, outputs a csv with the embeddings (Note! use for up to 100k embeddings)
4. python rag.py -> downloads gemma-2b-it, runs the RAG, and lets you have a chat with your PDF
5. (Optional) Run the dev versions (dev_preprocess_pdf.ipynb and dev_rag.ipynb) which uses llama-index as PDF reader and see the difference in the answer quality

PDF preprocessing

In the dev folder, I use the files for development, but also am using llama-index, at the time of writing using it requires an API key, which is free, but we do not know in the future~

In preprocess_pdf.py and rag.py I use just local, pip install and run libraries.

Demo of rag chat

rag_video_demo.mp4

Using out-of-the box embedding and language models

• embedding model: mixedbread-ai/mxbai-embed-large-v1
• LLM: google/gemma-2b-it

Any feedback is welcome! ^^