How muhammadzainrehmani/document-gpt works
Project type and purpose
• DocumentGPT is a full-stack, single-page Streamlit application for conversational document Q&A.
• It accepts user‐uploaded PDFs/DOCXs, chunks and embeds text, stores vectors in Qdrant, and routes queries through Gemini 1.5 Flash via LangChain.
File structure
• Top level:
• app.py – main Streamlit entrypoint and orchestration
• requirements.txt – Python dependencies
• PDF/ – sample document storage
• images/ – UI assets (if any)
• .streamlit/secrets.toml (described in README) – contains API keys for Gemini and Qdrant
• No separate “frontend” or “backend” folders; the app.py script houses both UI and server logic.
README for additional insights
• Embedding model: HuggingFace all-mpnet-base-v2
• Vector store: Qdrant (external service)
• LLM: Google Gemini 1.5 Flash via Generative AI API
• Pipeline steps: upload → chunk & metadata tag → embeddings → Qdrant upsert → chat interface → query embeddings → LLM call