Model Registry

capstone_agent/llm.py is the single place model selection happens. Every agent gets a Gemini built with HttpRetryOptions (exponential backoff) via build_model(tier) — never a bare model-id string.

Tiers

Tier	Model	Used by
`flash-lite` (default)	`gemini-3.1-flash-lite`	Orchestrator routing, validation, audit, simple stages
`pro`	`gemini-3.1-pro-preview`	Reasoning-heavy: SQL generation, answer synthesis, vision analysis
`pro-customtools`	`gemini-3.1-pro-preview-customtools`	Tool-heavy: evidence retrieval, query execution

The README also references a flash-image tier used by the Q&A and DB pipelines for generated visuals (image output).

Rules

Warning: Never bypass the registry

Never pass a bare model-id string to an ADK LlmAgent. Always llm.build_model(tier) — it adds retry/backoff and centralizes selection.

Add new tiers in llm.MODEL_TIERS; never invent model identifiers.

Gemini 3.1 on Vertex AI requires GOOGLE_CLOUD_LOCATION=global.

Auth paths (self-controlled live mode)

The public Cloud Run build is deterministic and does not receive model credentials. For a self-controlled live agent deployment, configure one of:

GOOGLE_API_KEY=<key> in .env
Vertex AI via ADC: gcloud auth application-default login + GOOGLE_GENAI_USE_VERTEXAI=TRUE + GOOGLE_CLOUD_PROJECT + GOOGLE_CLOUD_LOCATION=global

Then AGENT_EXECUTION_MODE=live. See Clinical App for how the product server switches modes.