Skip to content

README.md

Latest commit

 

History

History
76 lines (54 loc) · 3.66 KB

README.md

File metadata and controls

76 lines (54 loc) · 3.66 KB
Raw

Google GenAI Samples

These samples demonstrate the Temporal Google GenAI plugin, which runs the Google Gemini SDK inside Temporal Workflows. Workflows construct a TemporalAsyncClient, and every Gemini API call — generate_content, tool calls, streaming, files, interactions, agents — runs as a Temporal Activity. You get durable execution, Temporal-managed retries and timeouts, and your credentials never enter the workflow or its event history.

Samples

Sample Description
hello_world Minimal generate_content call. Start here.
tools Automatic function calling: an activity_as_tool-wrapped activity and a plain workflow-method tool on one call.
streaming Forward generate_content_stream chunks to an external subscriber via streaming_topic + WorkflowStream.
chat Multi-turn conversation with client.chats.
structured_output Typed JSON output via response_schema and a Pydantic model.
mcp Give Gemini an MCP server's tools via TemporalMcpClientSession.
files Upload a file with client.files and reference it in a call. (needs a live API key)
interactions Stateful server-side conversations via client.interactions. (needs a live API key)
agents Managed-agent CRUD via client.agents. (needs a live API key)
vertex_ai The hello-world flow against Vertex AI (vertexai=True). (needs GCP credentials)

Prerequisites

  1. Install dependencies:

    uv sync --group google-genai

  2. Configure credentials. Most samples use the Gemini Developer API and read an API key from the environment:

    export GOOGLE_API_KEY=...

    The vertex_ai sample instead uses Vertex AI with Google Cloud Application Default Credentials — see its README. You can authenticate with gcloud auth application-default login and set GOOGLE_CLOUD_PROJECT (and optionally GOOGLE_CLOUD_LOCATION).

  3. Start a Temporal dev server:

    temporal server start-dev

Running a Sample

Each sample has two scripts. Start the Worker first, then the Workflow starter in a separate terminal:

# Terminal 1: start the Worker
uv run google_genai/<sample>/run_worker.py

# Terminal 2: start the Workflow
uv run google_genai/<sample>/run_workflow.py

For example, to run the tools sample:

# Terminal 1
uv run google_genai/tools/run_worker.py

# Terminal 2
uv run google_genai/tools/run_workflow.py

Key Features Demonstrated

  • Durable API calls — every Gemini call runs as an activity with configurable timeouts and retries; no credentials enter workflow history.
  • Automatic function calling — the SDK's AFC loop runs in-workflow; tools can be durable activities (activity_as_tool) or plain workflow methods.
  • Streaming — forward model chunks live to external subscribers via WorkflowStream.
  • Structured output — Pydantic-typed results through the plugin's Pydantic data converter.
  • MCP integration — register MCP servers on the worker; tool calls dispatched through per-server activities.
  • Full API surface — chat, the Files API, the Interactions API, managed agents, and Vertex AI.

Related