RAG PDF QA System

A lightweight local PDF question-answering system built around the RAG (Retrieval-Augmented Generation) workflow. The project demonstrates how to turn unstructured PDF documents into a searchable knowledge base, retrieve relevant context with vector search, optionally rerank results, and generate final answers through an external LLM API.

Overview

This project is designed as a concise but complete RAG demo suitable for learning, portfolio presentation, and technical review. It focuses on the full retrieval pipeline rather than product-level complexity.

Core workflow:

1. Load local PDF files.
2. Split long text into smaller chunks.
3. Convert chunks into embeddings with a Hugging Face model.
4. Store and retrieve vectors with FAISS.
5. Optionally rerank retrieved chunks with a cross-encoder.
6. Assemble a grounded prompt from retrieved context.
7. Call an OpenAI-compatible external LLM API to generate the final answer.

Features

• Local PDF ingestion with PyPDFLoader
• Configurable chunking strategy
• Vector retrieval powered by FAISS
• Optional two-stage retrieval with reranking
• Streamlit chat-style interface
• External LLM integration through OpenAI-compatible API endpoints
• Environment-variable based API configuration
• Simple structure for fast understanding and demonstration

Tech Stack

• Python
• Streamlit
• LangChain
• FAISS
• Hugging Face Embeddings
• Sentence Transformers CrossEncoder
• PyPDF
• Requests

Retrieval Architecture

PDF files
  -> document loading
  -> text chunking
  -> embedding generation
  -> FAISS vector index
  -> similarity retrieval
  -> optional cross-encoder reranking
  -> prompt assembly
  -> external LLM answer generation

Project Structure

rag-pdf-qa-system/
├── app.py
├── requirements.txt
├── README.md
├── .gitignore
├── pdfs/                # put your local PDF files here
└── faiss_index/         # generated locally after building the vector index

Why This Project Is Strong for Review

• Demonstrates an end-to-end RAG pipeline rather than isolated API calls
• Separates retrieval and generation clearly, which makes the architecture easy to explain
• Uses practical open-source components commonly seen in real prototyping workflows
• Includes optional reranking, which shows awareness of retrieval quality optimization
• Keeps the implementation compact enough for reviewers to inspect quickly

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Prepare documents

Put one or more PDF files into the pdfs/ directory.

3. Configure optional API access

If you want the app to generate final answers through an external LLM, set these environment variables:

RAG_API_BASE=https://your-api-base.example/v1
RAG_API_KEY=your_api_key
RAG_MODEL_NAME=your_model_name

If these values are not configured, the system can still demonstrate retrieval and prompt assembly.

4. Run the app

streamlit run app.py

Usage

1. Start the Streamlit app.
2. Select a PDF from the sidebar.
3. Build or rebuild the vector index.
4. Ask questions in the chat input.
5. Inspect retrieved context blocks below the chat area.

Configurable Parameters

The sidebar supports adjusting the main retrieval parameters:

• Chunk Size
• Chunk Overlap
• Top-K
• Embedding model name
• Whether reranking is enabled
• Reranker model name
• Fetch-K for candidate recall

These controls make it easy to compare retrieval strategies during demos.

Typical Demo Questions

• What is the main topic of this PDF?
• Summarize the core functions described in the document.
• What installation or configuration steps are mentioned?
• What important parameters, models, or hardware details appear in the document?

Notes

• faiss_index/ is generated locally and should not be committed.
• pdfs/ is intentionally excluded from version control to avoid uploading private documents.
• API keys are never stored in the repository.
• This repository is intended as a clean public demo version.

Future Improvements

• Support multi-document indexing and source filtering
• Add file upload support in the UI
• Add source citation highlighting in answers
• Persist separate indexes for different document sets
• Introduce evaluation metrics for retrieval quality

License

This project is provided for learning, demonstration, and portfolio use.