How I Built InterSphere For the past few weeks, I’ve been building InterSphere, an AI-powered learning app that turns any topic into an explorable, visual “knowledge card.” Type Coffee, and it gives you a structured deep dive — definition, images, timeline, related concepts, and a quick quiz — all with citations. Here’s what’s happening under the hood: 1. Retrieval-Augmented Generation (RAG) Instead of letting the model invent answers, InterSphere first pulls real data from Wikipedia, Wikidata, and Openverse. That factual layer is passed to a lightweight LLM summarizer, which writes contextual explanations while preserving citations. Every paragraph in the app can trace back to a source. 2. Caching System LLM calls are expensive, so the API uses a hybrid cache: DiskCache for local storage (fast and simple). Redis layer for hot queries once traffic grows. Each knowledge card is cached for 7 days, and popular topics (like Coffee or AI ethics) are pre-computed weekly. Result: <300ms average load time and roughly 10× lower API cost. 3. The Connections Engine This part finds relationships between ideas — e.g., Coffee → Trade Routes → Ottoman Empire → Caffeine Science. It uses sentence-transformer embeddings stored in a pgvector database to compute semantic similarity. That’s how InterSphere builds its “map of knowledge.” Stack Summary: Next.js + Tailwind (frontend) FastAPI + Postgres + Redis (backend) RAG pipeline + caching + embeddings Why share this? Because I’m convinced that “search” shouldn’t just retrieve — it should teach. InterSphere is my small experiment in making information explorable, trustworthy, and alive. 🧩 Try it: https://intersphere.vercel.app 💬 Feedback: What other systems or topics should I connect next?

InterSphere

Every topic has a fascinating story

Every topic has a fascinating story