🔍 Retrieval-Augmented Generation (RAG) with Mistral + FAISS + PubMed

This project implements a Retrieval-Augmented Generation (RAG) pipeline using:

• 🧠 Mistral as the language model (generator)
• 📦 FAISS for vector-based document retrieval
• 📄 PubMed articles or any custom documents as external knowledge
• 🔗 Retrieval + Generation fused for factual, context-grounded responses

Author: Aromal Joseph 🚀 AIOps Engineer | LLM + RAG Enthusiast 📫 LinkedIn

---

📚 What is RAG?

Retrieval-Augmented Generation (RAG) enhances LLM responses by integrating external document retrieval into the generation process. Instead of relying solely on model parameters (parametric memory), RAG retrieves relevant text passages (non-parametric memory) and uses them as dynamic context during generation.

Memory Type	Description
Parametric	Model's internal knowledge (Mistral)
Non-Parametric	Retrieved documents (FAISS + PubMed)

---

💡 Use Cases

• Biomedical Q&A using PubMed abstracts
• Chatbots with custom knowledge bases
• Legal/financial assistants with grounded outputs
• Research support tools

---

🧱 Architecture Overview

+------------+     +---------------------+     +------------------------+
|  User Query| --> | Embed & Search FAISS| --> | Retrieve Top-k Chunks  |
+------------+     +---------------------+     +------------------------+
                                                            |
                                                            v
                                            +-------------------------------+
                                            | Mistral (Generator)           |
                                            | Prompt = Query + Retrieved    |
                                            +-------------------------------+
                                                            |
                                                            v
                                                +--------------------+
                                                | Final Answer Text  |
                                                +--------------------+

---

🧾 Dataset – PubMed Abstracts

You can use PubMed data (bio/medical literature) as your knowledge base.

🔽 To download sample PubMed abstracts:

wget https://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed23n0001.xml.gz
gunzip pubmed23n0001.xml.gz

Then parse it with your custom XML parser or a library like biopython.

---

⚙️ Step-by-Step Pipeline

1. Preprocess Documents

• Chunk .txt/.csv or PubMed abstracts into smaller passages (100–300 tokens).
• Clean stopwords, special chars.

2. Embed Chunks

• Use sentence-transformers for dense embeddings (e.g. all-MiniLM-L6-v2).

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = model.encode(text_chunks)

3. Create FAISS Index

import faiss
index = faiss.IndexFlatL2(embedding_dim)
index.add(np.array(embeddings))

Optional: Save the FAISS index for reuse:

faiss.write_index(index, "faiss_index/index.faiss")

4. Query Processing

query_embedding = model.encode([query])
_, top_k_indices = index.search(np.array(query_embedding), k=5)
retrieved_docs = [text_chunks[i] for i in top_k_indices[0]]

5. Prompt Construction

prompt = "Context:\n" + "\n".join(retrieved_docs) + f"\n\nQuestion: {query}\nAnswer:"

6. Answer Generation using Mistral

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")

inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
outputs = model.generate(**inputs, max_new_tokens=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

---

🏗️ Project Structure

rag_mistral/
├── data/                   # Raw documents or PubMed XMLs
├── embeddings/             # Saved embedding files
├── faiss_index/            # Serialized FAISS index
├── retriever.py            # Handles retrieval from FAISS
├── generator.py            # Loads Mistral and generates output
├── preprocess.py           # Cleaning, chunking, tokenizing
├── main.py                 # Glue code: pipeline from query to answer
├── requirements.txt        # All Python dependencies
└── README.md               # You are here

---

🧪 Example Command

python main.py --query "What causes type 1 diabetes?"

---

📦 Installation

✅ Python Dependencies

pip install faiss-cpu
pip install transformers
pip install sentence-transformers
pip install beautifulsoup4
pip install lxml
pip install tqdm

If using Mistral locally, install PyTorch (with GPU): https://pytorch.org/get-started/locally/

---

✍️ Sample Code Snippet (main.py)

from retriever import get_top_docs
from generator import generate_answer

query = "What is gene therapy?"
top_docs = get_top_docs(query)
answer = generate_answer(query, top_docs)

print("Answer:", answer)

---

🛠 Future Improvements

• Add BM25 retriever for hybrid search
• Optimize chunking strategy
• Add Streamlit/Gradio interface
• PubMed parser for full ingestion
• Caching and batching for speed

---

🧑‍💻 Author

Aromal Joseph AIOps Engineer | LLM Developer | Research + Product 🔗 LinkedIn

---

📜 License

MIT License — use, share, modify freely.

---

🙌 Acknowledgements

• FAISS – Facebook AI
• Hugging Face – Transformers & Mistral
• NCBI PubMed – Biomedical open data