Community Hub — Engineering Collective Landing Site

The site treats marketing as a performance problem: no content above the fold waits on a video, every scroll effect has a non-jacking fallback for mobile, and the AI assistant's retrieval index lives in module memory to avoid repeated embedding calls.

A single Next.js App Router page composed of independent sections. The chat widget uses a route handler that embeds the local knowledge corpus once per process, stores vectors in module memory, embeds each question as a RETRIEVAL_QUERY, ranks by cosine similarity, and passes the top 4 chunks as model context. The Gemini key is read only in the route handler and never reaches the browser.

A local knowledge corpus in src/lib/knowledge.ts — short, single-topic chunks — plus environment config for the Gemini API key and model name.

Balancing heavy scroll animation with mobile performance required fallbacks for every effect. The live credential builder had to render identically server and client. The RAG index needed to stay fast and memory-bounded for the life of the process.

Static rendering with dynamic chat only — the marketing content is served from the edge while the assistant route handler keeps a warm embedding index in memory.

Everything except the chat widget runs without a key; the widget returns a 500 and shows an inline error when the key is missing.

No custom telemetry — relies on platform analytics for the marketing page and error boundaries for the interactive sections.

HLS video hero over a static image for visual impact with progressive loading. Server-side RAG over client-side model calls to protect the API key. Module-memory vectors over a database for zero-infrastructure retrieval.