VertexConfig

VertexConfig defines all configurable parameters for the Google Vertex AI Gemini API, including authentication, function-calling modes, thinking configuration, and project-scoped routing.

Struct Definition

pub struct VertexConfig {
    pub api_key: Option<String>,
    pub access_token: Option<String>,
    pub base_url: String,
    pub model: String,
    pub project_id: Option<String>,
    pub location: String,
    pub stream: bool,
    pub max_output_tokens: Option<u32>,
    pub temperature: Option<f32>,
    pub top_p: Option<f32>,
    pub top_k: Option<u32>,
    pub function_calling_mode: VertexFunctionCallingMode,
    pub allowed_function_names: Option<Vec<String>>,
    pub stream_function_call_arguments: bool,
    pub thinking: Option<VertexThinkingConfig>,
    pub retry: Option<VertexRetryConfig>,
}

Fields

api_key

Optional API key for key-based authentication. Resolution order: config.api_key -> GOOGLE_VERTEX_API_KEY -> GOOGLE_API_KEY -> GEMINI_API_KEY.

access_token

Optional OAuth bearer token for Authorization header. Resolution: config.access_token -> GOOGLE_VERTEX_ACCESS_TOKEN.

base_url

Base URL for Vertex API requests. Defaults to https://aiplatform.googleapis.com.

model

Gemini model identifier. Defaults to "gemini-2.5-flash".

project_id

Google Cloud project ID. When set, project-scoped endpoints are used. Can be set via GOOGLE_VERTEX_PROJECT.

location

Vertex AI region for project-scoped endpoints. Defaults to "us-central1". Can be set via GOOGLE_VERTEX_LOCATION.

stream

Whether to call the streaming endpoint. Defaults to true.

max_output_tokens

Maximum output tokens. Defaults to 4096.

temperature

Sampling temperature (0.0--2.0). Validated during config validation.

top_p

Nucleus sampling (0.0--1.0). Validated during config validation.

top_k

Top-k sampling parameter.

function_calling_mode

Controls how the model handles function calling:

pub enum VertexFunctionCallingMode {
    Auto,  // Model decides (default)
    Any,   // Model must produce function calls
    None,  // Function calling disabled
}

allowed_function_names

Optional allow-list of function names when using Any mode.

stream_function_call_arguments

Whether to stream function call arguments as partial updates. Defaults to false.

thinking

Thinking configuration for supported Gemini models:

pub struct VertexThinkingConfig {
    pub thinking_level: Option<String>,  // "LOW", "MEDIUM", "HIGH"
    pub include_thoughts: Option<bool>,
}

• thinking_level -- Thinking level hint. Typical values: "LOW", "MEDIUM", "HIGH"
• include_thoughts -- Whether to include thought content in responses

retry

Retry policy for transient failures. Reuses OpenAI retry semantics for consistent behavior across providers. Default: 3 retries, 1s initial backoff, 60s max backoff.

Builder Methods

use appam::prelude::*;

let agent = AgentBuilder::new("my-agent")
    .provider(LlmProvider::Vertex)
    .model("gemini-2.5-flash")
    .max_tokens(8192)
    .system_prompt("You are a helpful assistant.")
    .build()?;

AgentBuilder does not expose a dedicated Vertex thinking helper today. Configure thinking through VertexConfig or [vertex.thinking] in TOML instead.

TOML Configuration

provider = "vertex"

[vertex]
model = "gemini-2.5-flash"
project_id = "my-gcp-project"
location = "us-central1"
max_output_tokens = 8192

[vertex.thinking]
thinking_level = "HIGH"
include_thoughts = true

Environment Variables

# API key authentication (checked in order)
export GOOGLE_VERTEX_API_KEY="..."
export GOOGLE_API_KEY="..."
export GEMINI_API_KEY="..."

# Or OAuth bearer token
export GOOGLE_VERTEX_ACCESS_TOKEN="ya29...."

# Project-scoped routing
export GOOGLE_VERTEX_PROJECT="my-gcp-project"
export GOOGLE_VERTEX_LOCATION="us-central1"

Validation

VertexConfig::validate() checks for:

• Empty model name (error)
• Empty location (error)
• Temperature outside [0.0, 2.0] (error)
• top_p outside [0.0, 1.0] (error)

Source

Defined in src/llm/vertex/config.rs.