Monitoring, testing, and troubleshooting AI solutions

Configuring and using Model Monitoring on Gemini Enterprise Agent Platform to establish continuous evaluation metrics for production models

Continuous Evaluation and Baseline Setup

Continuous evaluation is the ongoing process of measuring a deployed machine learning model's inference performance against ground truth labels and baseline datasets. To set up continuous evaluation, a model deployed to an endpoint logs incoming inference requests and responses directly to Cloud Storage or a BigQuery table. Vertex AI Model Monitoring then reads these logged production samples at scheduled intervals and compares them against the training baseline dataset. When ground truth labels become available from human reviewers or downstream business outcomes, the evaluation service calculates objective quality metrics like precision, recall, and root-mean-square error to verify that model quality remains stable over time.

Alerting and Automated Remediation Pipelines

Alerting and automated pipelines connect monitoring signals to operational responses when production metrics degrade. A monitoring job calculates drift and accuracy metrics, and if a metric crosses a preconfigured numerical threshold, it sends an alert through Cloud Monitoring to trigger notifications. The alert can also publish a message to a Cloud Pub/Sub topic that starts an automated workflow in Vertex AI Pipelines. The pipeline ingests the newly labeled production data, runs a model retraining job, validates the updated model against a held-out test dataset, and deploys the new model to the production endpoint only after it outperforms the active model.

Monitoring for common issues (e.g., training-serving skew, data drift, concept drift, and feature attribution drift)

Identifying Data Drift and Training-Serving Skew

Data drift and training-serving skew are distribution anomalies that occur when incoming production feature values differ from the data used during model development:

• Training-serving skew occurs immediately when the serving environment processes data differently than the training pipeline, such as through inconsistent preprocessing logic, differing feature store versions, or pipeline latency.
• Data drift occurs gradually when real-world distributions shift over time without changing the fundamental relationship between input features and target labels.

To detect these issues, Vertex AI Model Monitoring uses statistical distance metrics such as Jensen-Shannon divergence and Chebyshev distance to compare production feature distributions against baseline training distributions. When the calculated distance score exceeds the defined anomaly threshold, the system flags the specific feature that drifted.

Detecting Concept Drift and Feature Attribution Drift

Concept drift and feature attribution drift describe shifts in how inputs relate to outputs rather than simple changes in input values:

• Concept drift occurs when the statistical relationship between the input features and the target prediction label changes, causing an older model to make inaccurate predictions even if the input distributions appear normal.
• Feature attribution drift occurs when the relative importance or weight assigned to specific features shifts between the baseline dataset and live production requests.

Teams detect feature attribution drift by enabling Vertex Explainable AI on production endpoints, which computes feature importance scores (such as sampled Shapley values) for each incoming request. The monitoring service aggregates these individual attribution scores over a time window, compares the attribution rankings against baseline training explanations, and flags features whose explanatory influence has materially changed.

Monitoring, testing, and evaluating gen AI solutions

Automated and Model-Based Evaluation for Generative Models

Generative AI evaluation assesses large language model responses across quality, factual alignment, and task success using both heuristic and model-based scoring:

• Heuristic metrics compare generated text against gold-standard references using exact token overlaps, such as BLEU for translation precision and ROUGE for summarization recall.
• Model-based metrics use an evaluator foundation model to grade complex qualities such as coherence, fluency, task fulfillment, and grounding without requiring exact string matching.

In comparative evaluation workflows like AutoSxS, an evaluation pipeline sends identical prompt inputs to both a candidate model and a baseline model. The evaluator model scores both outputs side by side, determines a win-rate percentage, and generates natural language rationales explaining why one response outperformed the other.

Safety Filters and Production Guardrails

Safety monitoring and guardrails protect generative AI systems from harmful content, prompt injection attacks, and data leakage in production. Incoming user prompts first pass through safety filters that evaluate the input text across harm categories, including hate speech, harassment, dangerous content, and sexually explicit material. If a prompt or generated response violates the configured probability or severity threshold, the system blocks the content before it reaches the end user and returns a fallback response. Operations teams monitor production guardrail telemetry through aggregated audit logs in Cloud Logging to identify recurring prompt attacks, track blocked request rates, and adjust filter thresholds.

Key Facts

• Vertex AI Model Monitoring detects data drift by calculating statistical distance metrics, such as Jensen-Shannon divergence, between live production serving data and baseline training data.
• Training-serving skew occurs when differences in preprocessing pipelines, feature engineering code, or infrastructure cause serving features to diverge immediately from training features.
• Feature attribution drift uses Vertex Explainable AI to track changes in the relative importance scores of features over time.
• AutoSxS automates generative AI evaluation by using an evaluator model to compare outputs from two different models on the same prompt and generate win-rate statistics with explanations.
• Safety filters inspect generative inputs and outputs across standard harm categories, blocking content that exceeds configured severity thresholds before it reaches users.