Google Gemini

Google Gemini is an LLM plugin. It takes the transcribed conversation and generates the reply.

Setup

Set your Google API key in the worker environment. Generate a key from the Google AI Studio:

export GOOGLE_API_KEY=<key>

Usage

Import the plugin and pass it to the pipeline’s llm slot.

from zrt.plugins import GoogleLLM

llm = GoogleLLM(
    model="gemini-2.5-flash",
    thinking_budget=0,
    include_thoughts=False,
    max_output_tokens=8192,
)

# Pipeline(llm=llm, ...)

Vertex AI

By default the plugin calls the public Gemini API with GOOGLE_API_KEY. To run Gemini through Vertex AI instead, enable the Vertex backend and supply your Google Cloud project and location. Authenticate with a service account. Set GOOGLE_APPLICATION_CREDENTIALS (Python and JavaScript), or pass the key file path directly in Go:

export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json

from zrt.plugins import GoogleLLM

llm = GoogleLLM(
    model="gemini-2.5-flash",
    vertexai=True,
    project_id="your-gcp-project-id",
    location="us-central1",
)

# Pipeline(llm=llm, ...)

location defaults to us-central1 in all SDKs. In Python, project_id is required when vertexai=True and must be passed explicitly.

Configuration Options

Constructor parameters for the Python SDK (GoogleLLM). The JavaScript and Go SDKs expose equivalent options where supported, in their idiomatic form (camelCase / LLMOptions struct).

Core

Parameter	Type	Default	Description
`model`	`str`	`"gemini-2.5-flash-lite"`	Gemini model used to generate replies.
`api_key`	`str`	`None`	Google API key. Falls back to the `GOOGLE_API_KEY` environment variable when unset.
`temperature`	`float`	`0.7`	Sampling randomness; lower is more deterministic.
`max_output_tokens`	`int`	`8192`	Caps the length of each reply.
`top_p`	`float`	`None`	Nucleus-sampling probability cutoff.
`top_k`	`int`	`None`	Top-k sampling cutoff.
`presence_penalty`	`float`	`None`	Penalizes tokens already present, encouraging new topics.
`frequency_penalty`	`float`	`None`	Penalizes frequent tokens, reducing repetition.

Generation knobs

Parameter	Type	Default	Description
`seed`	`int`	`None`	Seed for reproducible sampling.

Vertex AI

Parameter	Type	Default	Description
`vertexai`	`bool`	`False`	Route requests through Vertex AI instead of the Gemini API.
`project_id`	`str`	`None`	Google Cloud project ID. Required when `vertexai` is enabled.
`location`	`str`	`"us-central1"`	Vertex AI regional endpoint used when `vertexai` is enabled.
`service_account_json`	`str \| dict`	`None`	Service-account credentials as a JSON string or dict, used for Vertex AI authentication.
`service_account_path`	`str`	`None`	Path to a service-account JSON key file. Falls back to the `GOOGLE_APPLICATION_CREDENTIALS` environment variable.

Safety

Parameter	Type	Default	Description
`safety_settings`	`list`	`None`	Content-safety thresholds.

Extended thinking

Parameter	Type	Default	Description
`thinking_budget`	`int`	`0`	Tokens the model may spend reasoning. `0` disables it for lower latency. Not available in the Go SDK.
`include_thoughts`	`bool`	`False`	Include the model’s reasoning in the output. Not available in the Go SDK.

For voice, short replies feel best. A zero or small thinking budget and a concise system instruction keep latency down.

Import paths

SDK	Import	Constructor
Python	`from zrt.plugins import GoogleLLM`	`GoogleLLM(...)`
JavaScript	`from '@zrt/js-sdk/plugins/google'`	`GoogleLLM({ ... })`
Go	`.../zrt-golang-sdk/plugins/google`	`google.NewLLM(google.LLMOptions{...})`

Plugins

Inference Gateway

Speech-to-text

LLM

Text-to-speech

Realtime models

Avatars

Turn detection

Voice activity detection

Noise cancellation

Setup

Usage

Vertex AI

Configuration Options

Core

Generation knobs

Vertex AI

Safety

Extended thinking

Import paths

​Setup

​Usage

​Vertex AI

​Configuration Options

​Core

​Generation knobs

​Vertex AI

​Safety

​Extended thinking

​Import paths

Setup

Usage

Vertex AI

Configuration Options

Core

Generation knobs

Vertex AI

Safety

Extended thinking

Import paths