📄 Retrieval-Augmented Generation (RAG) on PDF Documents

This project demonstrates a simple, end-to-end RAG pipeline using LangChain, Ollama, and FAISS to ask questions over a PDF document.

The pipeline:

1. Loads a PDF
2. Splits it into chunks
3. Embeds the chunks
4. Stores them in a vector database (FAISS)
5. Retrieves relevant chunks for a question
6. Uses an LLM to answer only based on the retrieved context

---

🧠 Model

• LLM: deepseek-r1:8b (via Ollama)
• Embeddings: Ollama embeddings using the same model

You can easily swap to another Ollama-supported model.

---

📁 Project Structure

.
├── duh.pdf                 # Input PDF document
├── rag_pdf.py              # Main RAG pipeline script
├── README.md               # Project documentation

---

⚙️ Requirements

Python

• Python 3.9+

Install Dependencies

pip install langchain langchain-community langchain-text-splitters \
            langchain-ollama faiss-cpu pypdf

Ollama

Install Ollama and pull the model:

ollama pull deepseek-r1:8b

Make sure Ollama is running:

ollama serve

---

🧩 Pipeline Overview

1️⃣ Load PDF

from langchain_community.document_loaders import PyPDFLoader

loader = PyPDFLoader("duh.pdf")
pages = loader.load()

---

2️⃣ Split into Chunks

from langchain_text_splitters import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=1500,
    chunk_overlap=100
)

chunks = splitter.split_documents(pages)

---

3️⃣ Create Embeddings & Vector Store

from langchain_community.vectorstores import FAISS
from langchain_ollama import OllamaEmbeddings

embeddings = OllamaEmbeddings(model=MODEL)
vectorstore = FAISS.from_documents(chunks, embeddings)
retriever = vectorstore.as_retriever()

---

4️⃣ Load LLM

from langchain_ollama import ChatOllama

model = ChatOllama(model=MODEL, temperature=0)

---

5️⃣ Prompt Template

The model is forced to answer only from context:

from langchain.prompts import PromptTemplate

template = """
You are an assistant that provides answers to questions based on
provided context.

If the answer is not in the context, reply: "I don't know".

Context:
{context}

Question:
{question}
"""

prompt = PromptTemplate.from_template(template)

---

6️⃣ Full RAG Chain

from operator import itemgetter
from langchain_core.output_parsers import StrOutputParser

chain = (
    {
        "context": itemgetter("question") | retriever,
        "question": itemgetter("question"),
    }
    | prompt
    | model
    | StrOutputParser()
)

---

❓ Asking Questions

questions = [
    "What can you get away with when you only have a small number of users?",
    "What's the most common unscalable thing founders have to do at the start?",
]

for q in questions:
    print(chain.invoke({"question": q}))

---

✅ Key Features

• 📄 PDF-based knowledge grounding
• 🔍 Semantic retrieval with FAISS
• 🧠 LLM answers constrained to context (low hallucination)
• 🔌 Runs fully locally (no cloud APIs)

---

🚀 Possible Extensions

• Add metadata filtering (page number, section)
• Persist FAISS index to disk
• Use multi-PDF ingestion
• Add citations / source highlighting
• Wrap with Streamlit or Gradio UI

---

🛡️ Hallucination Control

This project reduces hallucinations by:

• Strict context-based prompting
• Temperature = 0
• Retrieval before generation (RAG)

---

📌 Notes

• Designed for experimentation and learning
• Suitable for research papers, startup docs, or technical PDFs

---

🧑‍💻 Author

Avithal — Computer Vision & AI Engineer

---

⭐ If this helped you, consider starring the repo!