RAG Workshop: Building a PDF-Based Chatbot for Persian Documents

This repository contains the materials, notebooks, and code used in a hands-on workshop I presented titled "Development of an Offline Persian AI Assistant using RAG" at the WTLAB 2025 workshops. This workshop focused on building a Retrieval-Augmented Generation (RAG) chatbot for Persian PDF documents, such as legal regulations and organizational rules.

The workshop demonstrates how to extract text from PDFs, chunk and index the content, retrieve relevant documents, and generate final answers using open-source Large Language Models (LLMs). A simple Streamlit web interface is also included for interactive chat.

---

Workshop Goals

• Understand the basics of LLMs and Retrieval-Augmented Generation (RAG)
• Learn how to choose and use open-source models
• Learn how to parse Persian PDF documents efficiently
• Convert text to structured and searchable representations
• Implement a RAG pipeline to answer user questions from private domain documents with citations
• Deploy a simple chatbot UI using Streamlit

---

Repository Structure

File / Folder	Description
0.About-me.ipynb	Workshop introduction and presenter info
1.Introduction-to-LLMs-And-RAG.ipynb	Overview of LLMs and RAG concepts
2.GettingStarted-with-Open-Source-LLMs.ipynb	Practical intro to running and choosing open-source models
3.RAG-Workflow.ipynb	Explanation of RAG architecture
4.Parsing-PDFs.ipynb	Extracting and cleaning text from Persian PDFs
5.Implement_RAG.ipynb	Building a RAG system
8.More-Advanced-RAG.ipynb	Explanation of more advanced methods, Parent-Child Retriever, GraphRAG, etc.
7.streamlit_ui.py	Minimal chatbot UI built with Streamlit
data-sources/	Sample PDF regulation document used in the workshop
docling-pdf-parse/	Parsing PDF using docling
pyproject.toml, uv.lock	Environment configuration for reproducibility

---

Running the RAG Chatbot

1) Prepare the vector index (optional depending on notebook steps)

Run the relevant notebook to generate chunk embeddings.

2) Launch the Streamlit UI:

streamlit run 7.streamlit_ui.py

This opens a web interface where you can interact with your chatbot.

---

Notes on Persian PDF Processing

Persian PDFs are often encrypted, scanned, or improperly encoded. In this workshop, we compare multiple parsing approaches:

• PyPDF2
• OCR-based extraction
• Docling structured document parsing

Pros/cons and preprocessing techniques are discussed in the notebooks.

---

Future Extensions

You may extend this project by:

• Using FAISS / Qdrant / Weaviate for scalable vector search
• Implementing GraphRAG (entity/relation extraction → knowledge graphs) or Parent-Child Document Retriever

---

Author

Ali Moameri; Software Engineer & NLP Researcher

---