PDF RAG Pipeline

A local RAG (Retrieval-Augmented Generation) pipeline that indexes PDFs and answers questions about them.

Working Demo

Run the TF-IDF based demo (no external downloads needed):

python minimal_rag_demo.py

This demonstrates a complete RAG pipeline with your sample PDF!

Features

• PDF Text Extraction: Extracts and normalizes text from PDFs
• Smart Chunking: Splits text into overlapping chunks for better context
• Vector Embeddings: Multiple options (TF-IDF, SentenceTransformers)
• Similarity Search: Finds relevant content using cosine similarity
• Multiple Implementations: From simple to advanced

Quick Start

1. Basic Usage (Retrieval Only)

python pdf_qa.py --pdf your_document.pdf

This will:

• Extract and index the PDF
• Start an interactive Q&A session
• Show top matching passages for each question

2. With Local LLM (Answer Generation)

python pdf_qa.py --pdf your_document.pdf --use-llm

Uses a local language model to generate concise answers from retrieved passages.

Command Line Options

Option	Default	Description
--pdf	(required)	Path to your PDF file
--db	./pdf_index	ChromaDB persistence directory
--collection	pdf_qa	Collection name in ChromaDB
--reindex	False	Force re-indexing the PDF
--chunk	900	Chunk size in characters
--overlap	150	Overlap between chunks
--topk	4	Number of passages to retrieve
--use-llm	False	Use local LLM for answers
--model-id	microsoft/Phi-3-mini-4k-instruct	HuggingFace model ID

Examples

Index and query a PDF

python pdf_qa.py --pdf research_paper.pdf

Query with more context

python pdf_qa.py --pdf manual.pdf --topk 8

Force re-indexing

python pdf_qa.py --pdf document.pdf --reindex

Use with custom chunking

python pdf_qa.py --pdf large_doc.pdf --chunk 1500 --overlap 200

How It Works

1. Text Extraction: Reads PDF pages using pypdf
2. Chunking: Splits text into overlapping chunks for better context
3. Embedding: Converts chunks to vectors using all-MiniLM-L6-v2
4. Storage: Saves embeddings in ChromaDB (persistent)
5. Retrieval: Finds most relevant chunks using cosine similarity
6. Generation (optional): Uses local LLM to synthesize answers

Simple Example (mini_rag.py)

A minimal example showing the core concepts:

python mini_rag.py

This demonstrates RAG basics with tiny in-memory documents.

Tips

• First run downloads the embedding model (~80MB) - this is cached
• Using --use-llm downloads a larger model (~7GB for Phi-3)
• Indexed PDFs persist in the --db directory for reuse
• Adjust --chunk and --overlap for different document types
• Lower --topk for focused answers, higher for broader context

Requirements

See requirements in the code:

• pypdf
• sentence-transformers
• chromadb
• transformers
• torch