The Mind Surf
Full-stack multimodal RAG platform for chatting with PDFs
A full-stack, multimodal RAG platform that lets users upload PDFs — including ones packed with tables and images — and chat with them in real time, with hybrid search, reranking, and cited streaming answers.
Problem
General-purpose chatbots can't see a user's private documents, and most RAG tools fall apart on visually rich PDFs — tables and images get flattened or dropped during parsing, so the very content users most need to ask questions about becomes invisible to search. On top of that, naive RAG backends often block on CPU-heavy parsing/OCR work, freezing the app for every user whenever one person uploads a document.
Solution
RAG Compentator extracts text, HTML tables, and images from PDFs using Unstructured.io (with OCR support), then runs GPT-4o vision over table/image-bearing chunks to generate rich natural-language descriptions — turning previously "invisible" visual content into fully searchable text. Chunks are embedded into a hybrid dense+sparse Pinecone index, reranked with a hosted cross-encoder, and answers are generated with streaming citations grounded in the retrieved context. The whole pipeline is wrapped in a credit-gated, multi-user product with auth, namespaces, and an admin dashboard — deployed live at themindsurf.tech.
App Flow
Status
Live and deployed (themindsurf.tech), actively maintained with recent work on concurrent processing reliability and data-persistence hardening.