RAG PDF Chat

A web app built with Streamlit that lets you upload a PDF and ask questions about it in natural language. It uses Retrieval-Augmented Generation (RAG) to find the most relevant parts of the document and generate answers grounded in its content. If the answer isn’t in the PDF, it says so instead of making something up.

How it works

Users upload a PDF and chat through a Streamlit interface. Behind the scenes, the PDF is split into text chunks, converted into vector embeddings with Google Gemini and stored in ChromaDB. For each question, the most relevant chunks are retrieved by semantic similarity and sent to Gemini 3.6 Flash to generate the answer.

Challenges I solved

• Shared data between users: the first version used a single database on disk, so every visitor saw the last uploaded PDF. I refactored it to create an isolated in-memory database for each browser session.
• API rate limits: the Gemini free tier allows a limited number of requests per minute. I added page and chunk limits that validate the PDF before any API call, plus friendly error messages instead of crashes.
• Answer quality: I tuned chunk size and overlap and added a relevance threshold so the model only answers when it finds relevant context.

Tech stack

Python · LangChain · Google Gemini · ChromaDB · Streamlit · pytest