What's changed: Added per-section figures (cert-figure-retrofit). New AI-300 Chapter 4 (Domain 4 "GenAI quality assurance and observability": evaluation = test datasets/data mapping/AI quality metrics (groundedness/relevance/coherence/fluency)/risk-safety evaluation/automated evaluation workflows; observability = Foundry continuous monitoring/performance (latency/throughput/response time)/cost (token consumption/resource usage)/logging-tracing-debugging)
4.1Evaluating and validating generative AI apps and agents
Understand creating test datasets and data mapping, AI quality metrics (groundedness/relevance/coherence/fluency), risk and safety evaluations for harmful content detection, and configuring automated evaluation workflows using built-in/custom metrics.
Because generative AI output is non-deterministic, evaluation is central to operations. AI-300 asks you to evaluate quality and safety with measurable metrics and to automate it.
4.1.1Test datasets and quality metrics
For evaluation, prepare a test dataset (inputs and expected outputs/ground truth) and data mapping (defining column correspondence). Measure generative-AI quality with AI quality metrics: groundedness (is it based on the provided context = the opposite of hallucination), relevance (does it answer the question), coherence (is it logically consistent), and fluency (is it natural). For RAG, groundedness and relevance especially matter.
4.1.2Risk/safety evaluation and automated evaluation
Beyond quality, perform risk and safety evaluations: assess resistance to harmful content (violence, hate, sexual, self-harm) and jailbreaks, detecting unsafe output. Configure these as automated evaluation workflows, using built-in metrics plus custom metrics (your own criteria) as needed, evaluating continuously before and after deployment. The GenAIOps quality-assurance mindset: "make evaluation a pipeline and a quality gate."
| Aspect | Metric/means | Meaning |
|---|---|---|
| Grounded? | Groundedness | Opposite of hallucination |
| Answers the question? | Relevance | On-topic |
| Logically consistent? | Coherence | Logical flow |
| Natural language? | Fluency | Naturalness |
| Safe? | Risk/safety evaluation | Harmful content/jailbreak resistance |
| Evaluate continuously | Automated workflows (built-in/custom) | Quality gate |
Watch the mix-ups: (1) groundedness (based on provided context) vs relevance (answers the question) are different metrics—both matter for RAG. (2) quality metrics (correctness/naturalness) vs risk/safety evaluation (harmful content) have different goals. (3) Make evaluation an automated workflow for a continuous quality gate.
Map what to measure → metric: "measure ungrounded fabrication" = groundedness; "does it answer" = relevance; "logically consistent" = coherence; "natural language" = fluency; "detect harmful output" = risk/safety evaluation; "continuous quality gate" = automated evaluation workflows (built-in + custom).
4.1.3Section summary
- Evaluate with test datasets + data mapping; AI quality metrics = groundedness/relevance/coherence/fluency
- Risk/safety evaluation checks harmful content/jailbreak resistance
- Make automated evaluation workflows (built-in + custom) a quality gate
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which quality metric measures whether a response is based on the provided context (the opposite of hallucination)?
Q2. Which metric measures whether the response actually answers the question (on-topic)?
Q3. Which checks resistance to harmful content (violence, hate, sexual, self-harm) and jailbreaks?
Q4. What should you configure to make generative AI evaluation a continuous quality gate?
Q5. What do you prepare for generative AI evaluation that includes inputs and expected outputs (ground truth)?
Q6. Which correctly distinguishes groundedness from relevance?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.