Generative AI Google Brain

gemini-logging-intro-request-response-logging

3 min read

Intro to Request and Response Logging with Gemini

Source notebook

Repo path: gemini/logging/intro_request_response_logging.ipynb · Open on GitHub · intermediate

Shows how to enable Gemini request-response logging to BigQuery, query logs, and disable logging.

Summary

This notebook teaches how to configure Vertex AI request-response logging for a Gemini model using the preview Vertex AI SDK. The workflow initializes Vertex AI, enables logging with automatic BigQuery resource creation, sends a generate_content request, waits for logs, discovers the output_uri, queries recent log rows, and disables logging.

Key code patterns

Initialize Vertex AI

import vertexai
 
vertexai.init(project=PROJECT_ID, location=LOCATION)
model = GenerativeModel("gemini-2.5-flash")

Sets the Google Cloud project and region before creating the Gemini model client.

Enable logging

logging_config_result = model.set_request_response_logging_config(
    enabled=True,
    sampling_rate=1.0,
    bigquery_destination=f"bq://{PROJECT_ID}",
)
output_uri = logging_config_result.logging_config.bigquery_destination.output_uri

Turns on request-response logging and lets Vertex AI auto-create the BigQuery dataset and table.

Generate a logged request

response = model.generate_content("Why is the sky blue?")
print(response.text)
time.sleep(120)

A model call is needed to trigger BigQuery resource creation and produce a log entry.

Recover output URI

config = model.set_request_response_logging_config(
    enabled=True,
    sampling_rate=0.99,
    bigquery_destination=f"bq://{PROJECT_ID}",
)
output_uri = config.logging_config.bigquery_destination.output_uri

Uses a temporary config update to retrieve the auto-created BigQuery destination.

Query logs

uri_parts = output_uri.replace("bq://", "").split(".")
project_id, dataset_id, table_id = uri_parts
full_table_id = f"`{project_id}.{dataset_id}.{table_id}`"
df = bigquery.Client(project=project_id).query(query).to_dataframe()

Parses the BigQuery URI and reads recent logging records into a DataFrame.

Disable logging

model.set_request_response_logging_config(
    enabled=False,
    sampling_rate=1.0,
    bigquery_destination=f"bq://{PROJECT_ID}",
)

Uses the same configuration method to turn request-response logging off.

Models & APIs used

  • Models: gemini-2.5-flash
  • APIs / services: Vertex AI, BigQuery
  • SDKs / libraries: google-cloud-aiplatform, google-cloud-bigquery, vertexai

When to use this

Use this pattern when you need to audit, debug, or analyze sampled Gemini request and response payloads in BigQuery.

Gotchas & caveats

  • The logging feature is preview-only and uses vertexai.preview.generative_models.GenerativeModel.
  • The notebook requires a Google Cloud project ID and Vertex AI initialization in us-central1.
  • Colab users must authenticate with google.colab.auth.authenticate_user().
  • If logging is already enabled, set_request_response_logging_config can raise AlreadyExists.
  • BigQuery dataset, table, and log propagation can take a few minutes.
  • Automatically created datasets follow an ENDPOINT_DISPLAY_NAME_ENDPOINT_ID naming pattern.
  • Logged request-response pairs larger than the 10MB BigQuery write API row limit are not recorded.

Best practices

  • Use sampling_rate to control the fraction of requests logged.
  • Use an automatically created BigQuery destination by passing bq://PROJECT_ID.
  • Handle AlreadyExists when enabling logging.
  • Restore the original sampling_rate after temporarily updating the config to retrieve output_uri.
  • Quote the full BigQuery table ID before querying.
  • Disable request-response logging when it is no longer needed.

Graph View

Backlinks