eunki-7/Lightweight-PDF-RAG-Chatbot

DocuQuery AI – RAG Chatbot (Lightweight Fallback)

  

This project demonstrates a lightweight PDF Q&A chatbot using Retrieval-Augmented Generation (RAG). If no relevant content is found in the uploaded PDF, it automatically falls back to direct model inference.

Features

• —Document Upload & Parsing: Upload any PDF and extract text content
• —Embedding + Vector Search: Semantic search using all-MiniLM-L6-v2 + FAISS
• —RAG Pipeline: Answers with context from uploaded PDFs
• —Fallback Mode: Uses google/flan-t5-small for direct inference when no relevant context is found
• —Lightweight Deployment: Runs on Hugging Face Spaces free tier (CPU only)
• —Interactive Web UI built with Gradio

Architecture

[image]

• —PDF Loader: Extracts text from uploaded files
• —Vector Store: Embeds and stores document chunks
• —Retriever: Searches top relevant chunks
• —RAG Chain: Combines context with query
• —Fallback: Uses model-only generation when no context is relevant

Scalable AWS Deployment Architecture

If deployed on AWS, the architecture could be designed as follows:

• —API Gateway handles all external client requests.
• —AWS Lambda orchestrates:
• —Voice-to-Text using Amazon Transcribe or Whisper
• —Text-to-SQL or RAG inference via SageMaker Endpoints (FLAN-T5, MiniLM)
• —Amazon OpenSearch or Amazon RDS handles document retrieval or SQL execution.
• —Amazon S3 stores PDF documents and audio files.
• —Amazon CloudWatch monitors all components.

[image]

Production Considerations

Even though this is a free-tier prototype, the following practices are considered for production:

• —Security: IAM roles with least privilege, data encryption at rest and in transit
• —Scalability: Serverless (Lambda), containerized deployment, caching strategies
• —Monitoring: CloudWatch metrics and logging, automated alarms
• —CI/CD: GitHub Actions workflow for testing and automated deployment

Live Demo

**👉 Try it on Hugging Face Spaces**

Source Code

**📂 View on GitHub**

How It Works

1. 1.Upload a PDF document
2. 2.Ask a question in natural language
3. 3.The system searches for relevant document chunks:
4. 4.If context found → Answer based on PDF content
5. 5.If no context → Fallback to model-only response
6. 6.Returns both the answer and top context chunks

Tech Stack

• —Gradio – Web UI
• —Transformers – Model inference
• —FAISS – Vector similarity search
• —Python 3.10+

Installation & Run

pip install -r requirements.txt
python app.py

Example UI

[image]

• —This is the web-based interface built with Gradio

Use Cases

• —PDF-based Q&A systems (contracts, reports, manuals)
• —Knowledge base assistants
• —RAG + fallback strategy prototypes
• —Customer support document automation

Notes

• —First response may take a few seconds (model loading on CPU).
• —Subsequent responses are faster due to caching.

Author

eunki7

• —GitHub Profile