How to pass artifacts/files as on-demand context based on LLM decision to use the artifact on agent level?

[Satheesh-Balachandran]

I understand artifact service helps to load/save artifacts on the users/session scope.

But the use case I have is let's say I have different process files and based on the file names the LLM has to use the file contents only on demand, so context is not overloaded everytime when I use that agent?

What's the recommended approach towards building this use case?

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: The recommended approach is to use the Artifact Service to store your files and load them on-demand. There is a sample agent called "Sales Assistant Agent with Context Offloading" that demonstrates this exact pattern.

---

Hello! That's a great question. You are describing a pattern similar to Retrieval-Augmented Generation (RAG), where information is retrieved and provided to the LLM on-demand. ADK has excellent support for this.

The recommended approach is to use the Artifact Service [1]. This service is designed to manage large data blobs, like your process files. Instead of loading the entire file content into the context, you store the files as artifacts and then load them only when needed.

A perfect example for your use case is the "Sales Assistant Agent with Context Offloading" sample agent [2]. Here's how it works:

1. Save Artifacts: Large reports (in your case, your process files) are saved as artifacts using a custom tool.
2. Inform the LLM: The agent's instructions are updated with summaries of the available artifacts. This allows the LLM to know which files it can potentially use.
3. On-Demand Loading: The agent is equipped with a CustomLoadArtifactsTool. When the LLM determines that it needs the content of a specific file to answer a question, it uses this tool to load the file's content into the context for that specific turn.

This approach prevents the context from being overloaded with unnecessary information. The core functions you would use within a custom tool to achieve this are save_artifact and load_artifact [3].

For more advanced RAG capabilities, ADK also offers:

• RAG Agent Sample: This shows how to build a RAG agent using the Vertex AI RAG Engine for more complex scenarios.
• Spanner Tools RAG Agent Sample: This demonstrates building a RAG agent using Spanner's vector search capabilities [4].

I hope this helps you build your agent!

---

[1] https://google.github.io/adk-docs/artifacts/
[2] https://github.com/google/adk-python/blob/main/contributing/samples/context_offloading_with_artifact/README.md
[3] https://google.github.io/adk-docs/tools-custom/
[4] https://github.com/google/adk-python/blob/main/contributing/samples/spanner_rag_agent/README.md

[Satheesh-Balachandran]

But the artifacts are available only on user/session scope. I want to load artifacts on the agent level. Should I use tools + file path instead of Artifacts?