DocuMind AI: A Document-Powered Question Answering System Using Semantic Retrieval

Users often need answers from lengthy documents without knowing the exact terms used in the source. DocuMind AI is a document question-answering prototype that combines semantic retrieval with large language model generation. It accepts PDF, DOCX, TXT, and Markdown files, extracts and splits their text, creates embeddings using sentence-transformers/all-MiniLM-L6-v2, and indexes the chunks in FAISS. For each question, relevant chunks are supplied to ChatGroq using llama-3.3-70b-versatile. The Streamlit interface includes answer-length preferences, eight language options, recent conversation context, and dark/light modes. Project-reported preliminary evaluation found relevant content in the top three retrieved chunks for 87% of 30 factual questions and correct abstention on 19 of 20 out-of-context questions. These results come from small samples, and the test documentation is incomplete; they should not be treated as general performance guarantees. The prototype demonstrates a document-grounded interaction workflow while retaining limitations in extraction quality, retrieval coverage, multilingual reliability, and answer attribution.