MindStore AI
A retrieval-augmented generation (RAG) app — upload PDFs, text, or Markdown and chat with them using streamed AI answers grounded in your own documents, with local embeddings and no paid embedding API.
Overview
A retrieval-augmented generation (RAG) app — upload PDFs, text, or Markdown and chat with them using streamed AI answers grounded in your own documents, with local embeddings and no paid embedding API.
Tech Stack
Challenges & Learnings
MindStore AI answers questions from a user's own uploaded documents rather than a model's general knowledge. On upload, text is extracted, split into ~150-word overlapping chunks, and embedded locally with a Transformers.js pipeline (all-MiniLM-L6-v2) — avoiding per-call costs from a hosted embeddings API — then stored in Supabase Postgres. Retrieval is deliberately simple: rather than reaching for pgvector or an ANN index, every chunk is pulled back into memory and scored with a hand-rolled cosine-similarity function. That's a real tradeoff — easy to reason about and zero extra infrastructure, but it won't scale past a small document set before becoming a bottleneck. The top-matching chunks get stitched into a system prompt and streamed to the client through a hand-written SSE parser, rather than a higher-level SDK streaming helper. The app also fails gracefully: missing Supabase or AI provider config returns a clear 503 with a setup hint instead of crashing.