Runtime context_window_compression parameter #796 - 8

[adk-bot]

This pull request adds documentation for the new context_window_compression attribute in RunConfig, as requested in issue #796.

[joefernandez]

This change is not sufficient to document this feature. What are the settings of the ContextWindowCompressionConfig class? What's a minimal example of setting this config?

[jcwriter74]

Issue is closed given the below information

#19 Runtime context_window_compression parameter #807

Reasoning:

As part of the RunConfig, the ContextWindowCompressionConfig class serves as a way to manage the model’s context window. This feature helps by not exceeding the number of tokens of working memory of the model by compressing or sliding the model’s conversation history.

Implementation Example:

In a standard ADK agent.py or script using RunConfig, it is typically configured like this:

from google.adk.agents import RunConfig
from google.genai import types

run_config = RunConfig(
    context_window_compression=types.ContextWindowCompressionConfig(
        trigger_tokens=60000,
        sliding_window=types.SlidingWindow(target_tokens=5000)
    )
)

ContextWindowCompressionConfig class’ settings

The class’ settings are the following:

• trigger_tokens (int): describes a threshold amount of tokens after which the compression of the model’s conversation history mechanism is activated.

• sliding_window (types.SlidingWindow): defines how the context should be handled once triggered:
  ** target_tokens (int): specifies the number of tokens that should remain in the context after compression. The system will effectively "slide" the window, retaining only the most recent tokens up to this target amount.
  ** Example Usage: sliding_window=types.SlidingWindow(target_tokens=4000)