Document Q&A RAG System
Upload Documents, Get Instant AI Answers
Streamlit web application for intelligent document querying using Retrieval-Augmented Generation (RAG). Upload PDF or TXT files and ask natural language questions. Powered by Google Gemini embeddings, LangChain orchestration, and FAISS vector search. Includes multi-model support, chat history, and Docker deployment.
Why it mattered
Processes 10-50 page documents in 2-5 seconds; Vector-based similarity search with 4-chunk retrieval
What shipped
Streamlit web application for intelligent document querying using Retrieval-Augmented Generation (RAG). Upload PDF or TXT files and ask natural language questions. Powered by Google Gemini embeddings, LangChain orchestration, and FAISS vector search. Includes multi-model support, chat history, and Docker deployment.
Proof signal
Processes 10-50 page documents in 2-5 seconds; Vector-based similarity search with 4-chunk retrieval
System flow
Upload Document -> Extract Text -> Create Embeddings -> Index Vectors -> Retrieve Chunks -> Generate Answer.
Honest maturity snapshot
Raw datasets, disconnected metrics, and hard-to-scan findings.
A structured analysis or dashboard with clear KPIs, visual patterns, and decision-ready insights.
Features
- Streamlit integrated into the workflow.
- LangChain integrated into the workflow.
- Google Gemini integrated into the workflow.
- FAISS integrated into the workflow.
- Upload Document stage documented in the delivery flow.
- Extract Text stage documented in the delivery flow.
Engineering challenges
- Keeping the implementation clear enough to explain while still solving the core technical problem.
- Choosing a scope that could be shipped, tested, and documented.
- Turning technical work into proof a visitor can evaluate quickly.
Results
- Processes 10-50 page documents in 2-5 seconds; Vector-based similarity search with 4-chunk retrieval
- Documented the engineering path and important learnings.
- Made source code or technical proof available for review.
What I learned
- Tradeoff learned: Keeping the implementation clear enough to explain while still solving the core technical problem.
- Proof learned: Processes 10-50 page documents in 2-5 seconds; Vector-based similarity search with 4-chunk retrieval
- Next iteration: Add richer screenshots or a narrated demo.