Android LLM client architecture quiz

01

Product architecture

Why does the post argue that an Android AI feature should not be built as a thin wrapper around one model API?

A. Because all AI features should run only on device to avoid any networking code. Because the product should avoid every network call, every backend dependency, and every remote configuration flag.
B. Because real apps need routing, streaming state, memory, privacy, cost control, and failure handling around the model call.
C. Because Compose cannot render text that arrives from server-sent events.
D. Because remote providers always return identical streaming events and error states.

02

Inference routing

A. Always prefer remote inference because it has the largest models.
B. Always prefer local inference because it removes every privacy risk.
C. Route by task, data sensitivity, device state, network state, context size, cost, and provider health.
D. Ask the user to pick a model ID for every request before sending. and ask them to understand pricing, privacy, and provider capability details before continuing.

03

Provider layer

A. It hides the differences that matter, such as streaming events, tools, token usage, model capabilities, and rate limits.
B. It forces every model to run locally through AICore. and prevents any product team from choosing a cloud provider for high-quality multimodal tasks.
C. It makes the app store too much data in Room by default.
D. It prevents the app from showing any static answer after generation is done.

04

Streaming UI

A. Directly inside composables so the UI can parse every SSE chunk. including provider pings, partial tool deltas, retry headers, and transport-specific error payloads.
B. Inside the provider and data/domain layers, then reduced by the ViewModel into stable UI state.
C. Only in a background Worker, because streaming should never update the foreground UI.
D. Inside the navigation graph so every route can mutate the same text buffer.

05

Memory and context

A. So the app can always send the complete transcript to every provider. even if the transcript contains stale turns, private attachments, and content that exceeds the model window.
B. So every message becomes a permanent user preference.
C. So the app can pack context intentionally, respect privacy, control tokens, and support different model limits.
D. So local inference can skip storage and use only network state.

06

Privacy policy

A. Local inference solves all privacy problems by itself. and means logs, analytics, screenshots, attachments, and sync paths no longer need review.
B. Remote AI calls are data egress, and policy should decide what may leave the device before model routing happens.
C. Prompt privacy is only a server concern, not an Android client concern.
D. Crash logs and analytics cannot include AI prompt data.

07

Provider secrets

A. Because the keys can be extracted and abused, even if the app uses local encrypted storage.
B. Because Compose cannot call a backend proxy from a ViewModel. or from any screen that already has lifecycle-aware state collection and cancellation wiring.
C. Because remote providers reject every Android user agent.
D. Because Firebase AI Logic only supports server apps.

08

Token budgets

A. Reserving output tokens before context packing.
B. Using stable message IDs in a LazyColumn. and preserving scroll position while each partial assistant response arrives during generation.
C. Blindly retrying failed streams and sending huge transcripts without a budget.
D. Choosing a small local model for a low-risk rewrite task.

09

Hybrid routing

A. Because local models cannot stream any text on Android. regardless of whether the app uses AICore, MediaPipe, a downloaded model file, or another local runtime.
B. Because fallback may change the product promise if the remote path had tools, larger context, or stronger policy checks.
C. Because remote providers prevent apps from storing partial messages.
D. Because Android cannot detect whether the network is available.

10

Failure states

A. Pretend the answer completed successfully so the UI stays clean.
B. Delete the user message and force the user to type it again.
C. Save the partial assistant output with an interrupted state and offer clear retry or regenerate actions.
D. Retry every tool call automatically until the provider succeeds. including payment actions, external writes, and irreversible operations that already reached a backend.

11

On-device models

A. Only changing the button label from remote to local.
B. Model download UX, storage, RAM requirements, device gating, fallback, privacy copy, and failure handling.
C. Removing the data layer because local models do not need persistence. because the model can recreate every draft, attachment reference, preference, and previous message on demand.
D. Sending all prompts to analytics so the local model can improve.

12

Architecture summary

A. Render XML layouts, upload every prompt, and cache every provider response forever. while assuming privacy, routing, context packing, and cost control can all be handled later by the backend.
B. Feel like a normal mobile app, act like a privacy boundary, and act like a model router.
C. Avoid local storage, use one remote model, and hide all errors from users.
D. Replace the backend, replace the database, and replace user consent.