Enhance tool usage tracking with metadata aggregation #4424
Conversation
Summary of ChangesHello @jayy-77, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request significantly enhances the system's ability to track and report token usage at a granular level for individual tools and sub-agents. By introducing dedicated fields for tool usage metadata in event objects, providing utility for aggregating this data, and integrating it into the analytics and tracing infrastructure, the changes enable more precise cost analysis and performance monitoring for complex AI applications, especially those involving nested agent architectures and Vertex AI features. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
adk-bot
added
the
tracing
|
Response from ADK Triaging Agent Hello @jayy-77, thank you for creating this PR! To help reviewers understand and verify your changes, could you please fill out the "Testing Plan" section in your PR description? As per our contribution guidelines, this is required for all PRs that are not small documentation or typo fixes. Additionally, could you please clarify the relationship with issue #4419 or, if a dedicated issue doesn't exist for this change, describe the problem and solution in the PR description? This information will help the review process go more smoothly. Thanks! |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a valuable enhancement for tracking tool and sub-agent token usage by adding tool_usage_metadata. The implementation is well-integrated across the data models, telemetry, and agent logic. My review focuses on improving efficiency in the telemetry code, enhancing type safety for better maintainability, and adhering to Python's standard coding style.
| total_prompt_tokens = 0 | ||
| total_completion_tokens = 0 | ||
| total_tokens = 0 | ||
|
|
||
| for tool_name, usage_metadata in function_response_event.tool_usage_metadata.items(): | ||
| if usage_metadata: | ||
| total_prompt_tokens += getattr(usage_metadata, 'prompt_token_count', 0) or 0 | ||
| total_completion_tokens += getattr(usage_metadata, 'candidates_token_count', 0) or 0 | ||
| total_tokens += getattr(usage_metadata, 'total_token_count', 0) or 0 | ||
|
|
||
| if total_tokens > 0: | ||
| span.set_attribute(GEN_AI_USAGE_INPUT_TOKENS, total_prompt_tokens) | ||
| span.set_attribute(GEN_AI_USAGE_OUTPUT_TOKENS, total_completion_tokens) | ||
| span.set_attribute('gcp.vertex.agent.tool_usage_total_tokens', total_tokens) | ||
|
|
||
| # Add detailed breakdown as custom attribute | ||
| span.set_attribute( | ||
| 'gcp.vertex.agent.tool_usage_breakdown', | ||
| _safe_json_serialize({ | ||
| name: { | ||
| 'prompt_tokens': getattr(usage, 'prompt_token_count', 0) or 0, | ||
| 'completion_tokens': getattr(usage, 'candidates_token_count', 0) or 0, | ||
| 'total_tokens': getattr(usage, 'total_token_count', 0) or 0, | ||
| } | ||
| for name, usage in function_response_event.tool_usage_metadata.items() | ||
| }) | ||
| ) |
There was a problem hiding this comment.
The current implementation iterates over function_response_event.tool_usage_metadata twice: once to calculate the total token counts and again to create the breakdown dictionary. This can be optimized by performing both operations in a single loop. This refactoring will improve efficiency and make the code more concise.
total_prompt_tokens = 0
total_completion_tokens = 0
total_tokens = 0
breakdown = {}
for name, usage in function_response_event.tool_usage_metadata.items():
if not usage:
continue
prompt_tokens = getattr(usage, 'prompt_token_count', 0) or 0
completion_tokens = getattr(usage, 'candidates_token_count', 0) or 0
usage_total_tokens = getattr(usage, 'total_token_count', 0) or 0
total_prompt_tokens += prompt_tokens
total_completion_tokens += completion_tokens
total_tokens += usage_total_tokens
breakdown[name] = {
'prompt_tokens': prompt_tokens,
'completion_tokens': completion_tokens,
'total_tokens': usage_total_tokens,
}
if total_tokens > 0:
span.set_attribute(GEN_AI_USAGE_INPUT_TOKENS, total_prompt_tokens)
span.set_attribute(GEN_AI_USAGE_OUTPUT_TOKENS, total_completion_tokens)
span.set_attribute('gcp.vertex.agent.tool_usage_total_tokens', total_tokens)
# Add detailed breakdown as custom attribute
span.set_attribute(
'gcp.vertex.agent.tool_usage_breakdown',
_safe_json_serialize(breakdown),
)|
|
||
| # Aggregate and record usage for this sub-agent | ||
| if usage_metadata_list: | ||
| from ..models.llm_response import LlmResponse |
There was a problem hiding this comment.
This local import should be moved to the top of the file (e.g., with other from .. imports) to adhere to PEP 8 style guidelines and improve code readability. Top-level imports make dependencies explicit and easier to manage.
| super().__init__(invocation_context, event_actions=event_actions) | ||
| self.function_call_id = function_call_id | ||
| self.tool_confirmation = tool_confirmation | ||
| self._tool_usage: dict[str, Any] = {} |
There was a problem hiding this comment.
For better type safety, consider using a more specific type hint instead of Any. You can use a forward reference string (e.g., 'types.GenerateContentResponseUsageMetadata') to avoid circular import issues. This would also require adding from google.genai import types to the TYPE_CHECKING block at the top of the file.
| self._tool_usage: dict[str, Any] = {} | |
| self._tool_usage: dict[str, "types.GenerateContentResponseUsageMetadata"] = {} |
| def set_tool_usage( | ||
| self, | ||
| tool_name: str, | ||
| usage_metadata: Any, | ||
| ) -> None: | ||
| """Records usage metadata for a tool or sub-agent invocation. | ||
|
|
||
| Args: | ||
| tool_name: Name of the tool or agent that generated usage. | ||
| usage_metadata: Usage metadata object (GenerateContentResponseUsageMetadata | ||
| or dict with token counts). | ||
| """ | ||
| self._tool_usage[tool_name] = usage_metadata | ||
|
|
||
| def get_tool_usage(self, tool_name: str) -> Optional[Any]: | ||
| """Retrieves usage metadata for a specific tool. | ||
|
|
||
| Args: | ||
| tool_name: Name of the tool to retrieve usage for. | ||
|
|
||
| Returns: | ||
| Usage metadata if recorded, None otherwise. | ||
| """ | ||
| return self._tool_usage.get(tool_name) | ||
|
|
||
| def get_all_tool_usage(self) -> dict[str, Any]: | ||
| """Returns all tool usage metadata recorded in this context. | ||
|
|
||
| Returns: | ||
| Dictionary mapping tool names to their usage metadata. | ||
| """ | ||
| return self._tool_usage.copy() |
There was a problem hiding this comment.
To improve type safety and code clarity, the type hints for usage_metadata and the return types of these new methods should be made more specific than Any. Using 'types.GenerateContentResponseUsageMetadata' with a forward reference string is recommended. This change should be applied consistently with the update to _tool_usage.
def set_tool_usage(
self,
tool_name: str,
usage_metadata: "types.GenerateContentResponseUsageMetadata",
) -> None:
"""Records usage metadata for a tool or sub-agent invocation.
Args:
tool_name: Name of the tool or agent that generated usage.
usage_metadata: Usage metadata object (GenerateContentResponseUsageMetadata
or dict with token counts).
"""
self._tool_usage[tool_name] = usage_metadata
def get_tool_usage(self, tool_name: str) -> Optional["types.GenerateContentResponseUsageMetadata"]:
"""Retrieves usage metadata for a specific tool.
Args:
tool_name: Name of the tool to retrieve usage for.
Returns:
Usage metadata if recorded, None otherwise.
"""
return self._tool_usage.get(tool_name)
def get_all_tool_usage(self) -> dict[str, "types.GenerateContentResponseUsageMetadata"]:
"""Returns all tool usage metadata recorded in this context.
Returns:
Dictionary mapping tool names to their usage metadata.
"""
return self._tool_usage.copy()2 participants
Please ensure you have read the contribution guide before creating a pull request.
Link to Issue or Description of Change
1. Link to an existing issue (if applicable):
2. Or, if no issue exists, describe the change:
If applicable, please follow the issue templates to provide as much detail as
possible.
Problem:
A clear and concise description of what the problem is.
Solution:
A clear and concise description of what you want to happen and why you choose
this solution.
Testing Plan
Please describe the tests that you ran to verify your changes. This is required
for all PRs that are not small documentation or typo fixes.
Unit Tests:
Please include a summary of passed
pytestresults.Manual End-to-End (E2E) Tests:
Please provide instructions on how to manually test your changes, including any
necessary setup or configuration. Please provide logs or screenshots to help
reviewers better understand the fix.
Checklist
Additional context
Add any other context or screenshots about the feature request here.