Evaluation API: trajectory mismatch due to different JSON field order

[st4sik]

Hi, I’m testing my agent using the Evaluation API and have a question.

For example, I expect:

{
  "tool_name": "get_geo_coordinates",
  "tool_input": { "choice": "point", "location": "New York" }
}

But I get:

{
  "tool_name": "get_geo_coordinates",
  "tool_input": { "location": "New York", "choice": "point" }
}

The Evaluation API reports that the trajectories don’t match.

Am I correct that it treats this as different because of the field order? Shouldn’t it only be critical to have the same set of fields, regardless of their order?

[adk-bot]

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

TLDR: You are correct. The Evaluation API currently considers the order of fields in the JSON object, which is why you are seeing a mismatch.

---

Hi, thanks for reaching out and for your detailed example.

You are correct in your analysis. The current implementation of the Evaluation API compares the tool calls in a way that is sensitive to the order of fields in the tool_input JSON object. This means that even if the same fields and values are present, a different order will result in the trajectories not matching.

While the trajectory evaluator is designed to ignore extra keys in the top-level tool call dictionary (like keys other than "tool_name" and "tool_input"), it appears the comparison of the nested tool_input object is not canonicalized to account for different field orders [1].

This is a great point, and your feedback is valuable. For now, to ensure your trajectories match, you would need to ensure that the field order in your agent's output matches the expected order in your test cases.

We appreciate you bringing this to our attention.

---

[1] test_trajectory_evaluator.py

[st4sik]

It would be great to get an official response from a developer.