abdurrafey237/rag-chatbot

This is a document question-answering app called Citera. You upload your own PDF, Word, text, or CSV files, ask questions in plain language, and the app answers using only the content from those files, with the source passages shown alongside each answer. If your question falls outside what your documents cover, the app says so rather than guessing. This is useful for reading research papers, legal documents, company reports, or any collection of files where you want to find specific information without reading everything yourself. The key idea is that AI models trained on general knowledge often invent plausible-sounding but wrong answers when asked about private documents or recent information they were never trained on. This app works differently: it takes your question, finds the most relevant passages from your uploaded files using a vector search, and passes only those passages to the AI model for the answer. The answer is grounded in what your files actually say. You can choose from three AI providers (OpenAI, Google Gemini, and Hugging Face), switch between multiple retrieval strategies including an optional reranker that sharpens which passages are used, and get responses in ten languages. You can also build a document index once and reopen it in future sessions without re-uploading, which speeds up repeated use of the same files. The app runs in a browser-based interface built with Streamlit. Getting it running locally requires Python 3.11 specifically (the dependencies do not install on newer Python versions), a virtual environment setup, and an API key from whichever provider you choose. Google Gemini is the recommended starting point because a single Gemini key covers both the document indexing step and the answer generation step.