A retrieval-augmented generation (RAG) demo over
iPullRank's AI Search Manual.
The 23 chapters are split into text chunks and viewport-sized screenshot tiles, embedded with
gemini-embedding-2 into a Chroma index:
348 text chunks + 455 screenshot tiles = 803 vectors.
Each question retrieves the closest matches; gemini-3.5-flash-lite streams an answer grounded only in that context.
The API is self-hosted behind a Cloudflare Tunnel; this page is a same-origin proxy that forwards requests and streams tokens back over SSE. Free queries are capped per session and per IP. Use Step-through to watch retrieval before generation. Retrieved screenshot tiles appear under each answer; if the model can't answer, a Why no answer? panel shows what was retrieved and how closely it matched.