A full-stack document question-answering platform that allows users to upload PDFs or text files and ask natural-language questions using a retrieval-augmented generation (RAG) pipeline, with real-time responses and transparent source context.
Large language models are limited when working with long-form documents or private data without an external retrieval mechanism.
The goal of this project was to design an end-to-end document Q&A system that allows users to query their own files accurately, efficiently, and transparently.
The application needed to support real-time responses, reliable document retrieval, strong test coverage, and a clean, intuitive user experience.
I approached this as a production-style system rather than a simple LangChain demo.
Instead of treating document ingestion and question answering as isolated steps, I designed a complete retrieval-augmented generation (RAG) pipeline with clear separation of concerns.
A strong emphasis was placed on observability, testability, and frontend-backend communication to ensure the system could be extended and debugged effectively.
Designed a FastAPI backend to handle document uploads, ingestion, vector storage, and question-answering endpoints
Implemented document parsing, chunking, and embedding using LangChain and OpenAI embeddings
Stored document vectors in a local ChromaDB vector database for fast similarity search
Built a LangChain QA chain to retrieve relevant document chunks and generate answers via an LLM
Streamed LLM responses token-by-token back to the frontend for real-time interaction
Developed a React + Vite frontend with Tailwind CSS for drag-and-drop uploads and question submission
Added a vector store management interface to reset and manage stored documents during testing
Implemented a full backend test suite using pytest, httpx, and pytest-asyncio
This project emphasized reliability, transparency, and controlled data flow rather than traditional application security flaws.
Clear boundaries were enforced between user-uploaded content, vector storage, and LLM interaction.
LangSmith tracing was integrated to provide full visibility into chain execution, LLM calls, inputs, outputs, and performance timing.
Environment-based configuration ensures sensitive API keys and tracing controls are never hardcoded into the application.
Successfully built a fully functional document Q&A platform with real-time, streamed LLM responses.
Implemented transparent source attribution with confidence-based color coding to improve trust and interpretability.
Achieved strong code quality with a Pylint score of 9.77/10 and zero known linting errors at time of completion.
Delivered a scalable foundation for future AI-powered document analysis features such as multi-file querying, authentication, and cloud deployment.
This project is fully documented on GitHub, including notes, commits, and future ideas.
