Separate fast path from heavy path
Some requests should stay cheap and fast; others can invoke tools, retrieval, or larger multimodal flows.
Free Gemini breakdown
How to answer Design Gemini in a system design interview.
Strong answers treat Gemini-like systems as orchestration problems across models, tools, memory, and latency, not as “one LLM behind an API.”
Center of gravityOrchestration under latency pressure.
Multimodal inputTool routingFast responses
The pivot
The system is more than a model call.
Treat tools as first-class
Grounded assistant behavior depends on when to call tools, not just how good the model is.
Budget latency
Latency is a product feature in assistants. Retrieval, memory, safety, and multimodal processing all compete for it.
Want the full version?
The paid Gemini book goes deeper on assistant architecture.
The full breakdown covers multimodal input, memory, tool use, orchestration, latency budgeting, and productized assistant tradeoffs.