This project is an AI-powered PDF Assistant that allows users to ask questions about the contents of a PDF document and receive accurate answers. The system uses Retrieval-Augmented Generation (RAG), which combines semantic search with a language model to generate context-aware answers.

The application extracts text from a PDF, splits it into smaller chunks, converts those chunks into vector embeddings, stores them in a FAISS vector database, retrieves the most relevant chunks for a query, and then generates a final answer using a language model.

• Enable users to ask questions about a PDF document
• Retrieve the most relevant information from the document
• Generate clear and complete answers
• Provide an automatic summary of the entire document
• Package the system using Docker for easy deployment

The system follows a Retrieval-Augmented Generation (RAG) pipeline:

1. PDF Input

   • The user uploads a PDF file.

2. Text Extraction

   • Text is extracted from the PDF using PyMuPDF.

3. Text Chunking

   • The document is split into smaller chunks to improve retrieval accuracy.

4. Embedding Generation

   • Each chunk is converted into a numerical vector representation using a sentence transformer model.

5. Vector Storage

   • All embeddings are stored in a FAISS vector index for fast similarity search.

6. Query Processing

   • When a user asks a question, the system converts the question into an embedding.

7. Semantic Retrieval

   • FAISS retrieves the most relevant text chunks based on similarity.

8. Answer Generation

   • The retrieved context and user question are given to a language model to generate the final answer.

9. Document Summarization

   • The system also summarizes each chunk and combines them into a final summary of the whole document.

Used as the main programming language for building the entire pipeline.

Used for extracting text from PDF documents.

Purpose:

• Read PDF files
• Extract text page by page

Used for splitting the extracted text into sentences.

Purpose:

• Sentence tokenization
• Preparing text for chunking

Model used:

sentence-transformers/all-mpnet-base-v2

Purpose:

• Convert text chunks into semantic embeddings
• Enable similarity search between queries and document chunks

Purpose:

• Store embeddings efficiently
• Perform fast similarity search
• Retrieve the most relevant document chunks for a query

Model used:

google/flan-t5-large

Purpose:

• Generate answers based on retrieved context
• Summarize document chunks
• Produce natural language responses

Used for:

• Handling embedding vectors
• Converting embeddings into arrays for FAISS indexing

Used to build the API server that allows the system to be accessed through HTTP requests.

Purpose:

• Serve the AI model as an API
• Allow users to send queries and receive answers

Used to containerize the application.

Purpose:

• Ensure the application runs consistently on any system
• Package dependencies, models, and code together
• Simplify deployment

Used as the main development environment for running and managing the project locally.

PDF File
   ↓
Text Extraction (PyMuPDF)
   ↓
Sentence Splitting (Regex)
   ↓
Chunk Creation
   ↓
Embedding Generation (Sentence Transformers)
   ↓
Vector Storage (FAISS)
   ↓
User Question
   ↓
Query Embedding
   ↓
Relevant Chunk Retrieval
   ↓
Answer Generation (FLAN-T5)

docker build -t pdfasst .

docker run -p 8000:8000 pdfasst

Open in browser:

http://localhost:8000/docs

This opens the FastAPI Swagger interface where users can interact with the API.

What is cryptography?

The system retrieves relevant information from the PDF and generates a clear answer using the language model.

The system also generates a complete summary of the PDF by:

1. Summarizing individual chunks
2. Combining them into a final overall summary

• Works with one PDF at a time
• Large models require more memory
• First run may take time due to model loading

Tharun P

Artificial Intelligence / Retrieval-Augmented Generation / Natural Language Processing