Perplexity

Build me a personal answer engine like Perplexity, wrapping real search and LLM APIs, not rebuilding them. Requirements: - A local web app on localhost:3000: Node + Express + htmx, one input box, streamed answers. - Per question: query the Brave Search API (key in .env), fetch the top 6 results with undici, extract readable text with @mozilla/readability + jsdom, then pass the question plus numbered excerpts to Claude or GPT (key in .env) with instructions to answer only from the excerpts and cite as [1][2]. - Render citations as footnote links; show the full source list under every answer. - Follow-up questions stay in the same thread with prior Q&A in the context window. - Threads stored in SQLite via better-sqlite3; a sidebar lists past threads by first question. - If a page fails to fetch or extract, drop it and continue. Never cite a page that was not fetched. - Binds to localhost only; no accounts, no telemetry, only the search and LLM calls leave my machine. - Out of scope: crawling my own web index, model routing, and mobile apps. Answer quality tracks the search API, that is the deal. - README: which keys to get (Brave Search has a free tier) and rough per-query cost.