Gemini Interactions API

Examples using Google's Interactions API with Agno.

The Interactions API is a new primitive that provides:

Server-side conversation history - Only send new messages each turn, not the full history
Implicit caching - Prior turns are cached server-side for lower costs and latency
Typed execution steps - Responses contain discriminated content types for better observability
Background execution - Support for long-running tasks
Multimodal I/O - Image, audio, video, and document inputs; image and audio generation

Setup

bash

pip install -U google-genai
export GOOGLE_API_KEY=your-api-key

Requires google-genai>=2.0.0.

Examples

File	Description
`basic.py`	Basic text generation (sync, async, streaming)
`tool_use.py`	Function calling with external tools
`multi_turn.py`	Multi-turn conversation with server-side history
`thinking.py`	Reasoning/thinking mode
`search.py`	Built-in Google Search tool
`image_understanding.py`	Image analysis from URLs, files, and bytes
`image_generation.py`	Generate images with response_modalities
`audio_understanding.py`	Audio analysis and transcription
`video_understanding.py`	Video analysis from URLs
`document_processing.py`	PDF document processing
`structured_output.py`	Structured JSON output with Pydantic schemas

Usage

python

from agno.agent import Agent
from agno.models.google import GeminiInteractions

agent = Agent(
    model=GeminiInteractions(id="gemini-3-flash-preview"),
    markdown=True,
)
agent.print_response("Hello!")

Image Understanding

python

from agno.media import Image

agent.print_response(
    "What is in this image?",
    images=[Image(url="https://example.com/photo.jpg")],
)

Structured Output

python

from pydantic import BaseModel

class MovieReview(BaseModel):
    title: str
    rating: float

agent = Agent(
    model=GeminiInteractions(id="gemini-3-flash-preview"),
    output_schema=MovieReview,
)

Inference Tiers

python

# Lower cost, higher latency
agent = Agent(
    model=GeminiInteractions(id="gemini-3-flash-preview", service_tier="flex"),
)

# Lowest latency
agent = Agent(
    model=GeminiInteractions(id="gemini-3-flash-preview", service_tier="priority"),
)

Notes

The Interactions API is experimental and may change in future versions
Interactions are stored server-side for 55 days (paid) / 1 day (free tier)
System instructions and tools must be re-sent each turn (they are interaction-scoped)
Set store=False to disable server-side persistence