Artifact Implementation

[TooGoodDT]

I'm a bit confused as to the implementation requirements of artifacts. If I want to implement them is it required that I set up my own custom runner, artifact service, memory service, etc. or can I integrate within the default structure in which we are only required to define our agent and tools?

If it is possible to implement without requiring the creation of a main.py to run everything and can be utilized through adk web, how do we add this feature correctly? I want my agent to be able to access a large store of information whenever a user query necessitates that the agent reference it but do so without flooding its context.

Any shared code on the most simple method of saving and loading artifacts correctly so that the agent can interact with them would be much appreciated, thanks!

[agammann]

With current Python ADK, you can keep the normal my_agent/agent.py + adk web setup. The web launcher supplies a local artifact service by default (under the agent's .adk/artifacts directory), so you do not need a separate main.py or custom Runner just to use artifacts.

For a small named text artifact, add tools like these to your existing agent:

from google.adk.tools.tool_context import ToolContext
from google.genai import types

async def save_note(text: str, tool_context: ToolContext) -> str:
    part = types.Part.from_bytes(data=text.encode("utf-8"), mime_type="text/plain")
    version = await tool_context.save_artifact(filename="user:note.txt", artifact=part)
    return f"Saved version {version}"

async def load_note(tool_context: ToolContext) -> str:
    part = await tool_context.load_artifact(filename="user:note.txt")
    if part is None or part.inline_data is None:
        return "No note has been saved."
    return part.inline_data.data.decode("utf-8")

Include save_note and load_note in your agent's tools=[...] list. The artifact guide covers these async context methods. The user: prefix makes the note available across that user's sessions; omit it for a session-only file.

For a large searchable information store, artifacts alone are not an index: load_artifact retrieves a named file. Put the documents in a retrieval system and give the agent a search tool that returns only relevant snippets for each question (see the memory/retrieval guide). That avoids putting the whole store into every prompt.