AICore Multimodal Prompt Orchestration Script Generator

Role: Lead Android Systems Architect specializing in Google AICore and edge generative foundation models.

Context

• AICore integration level: {{aicore_client_version}}
• Supported input stream types: {{multimodal_input_modalities}}
• Maximum inference context window: {{token_budget_per_turn}}
• Streaming response timeout: {{stream_timeout_ms}}
• Fault tolerance policy: {{error_retry_strategy}}
• UI binding pattern: {{compose_state_architecture}}

Task

Author a production-grade Kotlin script implementing an asynchronous Repository and ViewModel architecture that consumes multimodal prompt inputs (bitmaps, audio descriptors, and text) to orchestrate streaming generative completions through Android AICore, emitting unidirectional UI state.

Method

1. Configure the AICore model binding and availability checks, ensuring compatibility with {{aicore_client_version}}.
2. Construct an input preprocessor that validates, scales, and tokenizes {{multimodal_input_modalities}} within {{token_budget_per_turn}} limits.
3. Build an asynchronous streaming engine using Kotlin Flow<T> that consumes generative token chunks from AICore without blocking.
4. Implement backpressure handling to ensure streaming UI updates conform to {{compose_state_architecture}}.
5. Enforce structured concurrency with CoroutineScopes linked to Android ViewModel lifecycles.
6. Integrate an automated recovery handler executing {{error_retry_strategy}} upon model eviction or timeout.
7. Wrap the generation lifecycle in an instrumentation layer measuring Time-To-First-Token (TTFT) and token generation velocity.

Constraints

• Code MUST use Kotlin Flow for streaming generation and avoid standard blocking RPC calls.
• MUST NOT exceed {{token_budget_per_turn}} under any input truncation path.
• Sensitive prompt payloads MUST be cleared from memory buffers immediately after dispatch.
• State management MUST strictly adhere to Unidirectional Data Flow (UDF).

Output format

1. Dependency Setup: Build configuration specifications.
2. Domain Models & Contracts: Sealed classes for MultimodalInputs, GenerationState, and Errors.
3. AICore Repository Script: Complete streaming client with backpressure management.
4. ViewModel State Engine: Full implementation binding to {{compose_state_architecture}}.

Self-review

• Verify all Flow emissions are thread-confined and lifecycle-aware.
• Ensure non-image tokens and image bytebuffers are recycled after model ingestion.
• Confirm the timeout mechanism fires at {{stream_timeout_ms}} without leaking active coroutines.