Struct GeminiLlm 

pub struct GeminiLlm { /* private fields */ }

Concrete Gemini LLM implementation using gemini-live Client.

The gemini-live Client is created once at construction time and reused for all generate() calls, matching the JS GenAI SDK pattern where a single GoogleGenAI instance is shared across requests.

Implementations

impl GeminiLlm

pub fn new(params: GeminiLlmParams) -> Self

Create a new GeminiLlm from parameters.

Resolves defaults for model, variant, API key, project, and location from parameters first, then falls back to environment variables. The gemini-live Client is created once here and reused for all calls.

pub fn from_env() -> Result<Self, LlmError>

A GeminiLlm configured from the environment, checked before any request is made.

Reads the same variables as new — GEMINI_API_KEY (or GOOGLE_GENAI_API_KEY, GOOGLE_API_KEY) for Google AI; GOOGLE_GENAI_USE_VERTEXAI=true with GOOGLE_CLOUD_PROJECT and optionally GOOGLE_CLOUD_LOCATION for Vertex AI; GEMINI_TEXT_MODEL or GEMINI_MODEL for the model — and fails with a message naming the variable to set when one is missing, or when the model is a Live model the text API does not serve. new accepts the same configuration and fails on the first request instead.

use gemini_adk_rs::llm::GeminiLlm;

let llm = GeminiLlm::from_env()?;

pub fn try_new(params: GeminiLlmParams) -> Result<Self, LlmError>

Like new, but checks the resolved configuration first; see from_env.

pub fn is_supported(model: &str) -> bool

Check if a model name is supported by GeminiLlm.

pub fn variant(&self) -> GoogleLlmVariant

Get the variant (VertexAI vs GeminiApi).

Trait Implementations

impl BaseLlm for GeminiLlm

fn warm_up<'life0, 'async_trait>( &'life0 self, ) -> Pin<Box<dyn Future<Output = Result<(), LlmError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Pre-warm the HTTP connection pool by making a lightweight request.

Establishes the TCP+TLS connection so the first real generate() call doesn’t pay the ~100-300ms handshake penalty. reqwest’s connection pool keeps it alive for subsequent calls.

fn model_id(&self) -> &str

The model identifier (e.g., “gemini-2.5-flash”).

fn generate<'life0, 'async_trait>( &'life0 self, request: LlmRequest, ) -> Pin<Box<dyn Future<Output = Result<LlmResponse, LlmError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Generate content from the LLM.

fn generate_stream<'life0, 'async_trait>( &'life0 self, request: LlmRequest, ) -> Pin<Box<dyn Future<Output = Result<LlmStream, LlmError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Generate content as a stream of chunks, in the order the model produced them. LlmResponse::append folds them into the whole reply; usage, when a chunk carries it, is cumulative.

fn capabilities(&self) -> ModelCapabilities

What this model supports — the ADK model-capability-declaration pattern: the model states its capabilities instead of every caller re-inferring them from the id string. The default derives a conservative estimate from model_id; back-ends with authoritative knowledge should override.