Nutrition RAG Pipeline (Gemma 2B + MPNet)

This repo contains a from-scratch RAG pipeline built in Google Colab, using:

• A 1200+ page human nutrition PDF as the knowledge source
• Sentence-based chunking (10 sentences per chunk)
• all-mpnet-base-v2 embeddings from Sentence-Transformers
• A local instruction-tuned LLM (google/gemma-2b-it) loaded with 4-bit quantization
• A retrieval → augmentation → generation loop closely following the “Production Level RAG Workshop (Part 2)” by Vizuara

The goal is to show the full RAG workflow without LangChain / LlamaIndex.

---

Pipeline Overview

1. Ingestion

   • Load the Human Nutrition PDF with PyMuPDF
   • Clean page text (remove newlines, extra spaces)
   • Optionally skip the first 41 pages (front matter)

2. Chunking

   • Use spacy’s English() + sentencizer to split each page into sentences
   • Group sentences into chunks of 10 sentences
   • Store for each chunk: page number, chunk index, and sentence text (sentence_chunk)

3. Embedding

   • Load all-mpnet-base-v2 from sentence-transformers
   • Encode all chunks into 768-dim embeddings
   • Keep embeddings as a torch tensor on GPU for fast retrieval

4. Retrieval

   • For a user query:
     • Embed the query with the same MPNet model
     • Compute dot-product (cosine) similarity between query and all chunk embeddings
     • Take top-k chunks as context

5. Augmentation (Prompt Formatting)

   • Combine the top-k chunks into a bullet list
   • Insert them into a long instruction prompt with a few example Q&A pairs
   • Wrap the final text using tokenizer.apply_chat_template(...) for Gemma

6. Generation (Local LLM)

   • Load google/gemma-2b-it with:
     • 4-bit quantization via bitsandbytes
     • flash_attention_2 when available, otherwise sdpa
   • Tokenize the prompt, send to GPU, and call generate(...)
   • Decode the output and strip the original prompt to get the final RAG answer

---

Setup

1. Clone the repo

git clone https://github.com/sgundala/RAG-from-Scratch.git
cd rag-nutrition-rag