Governance and Evaluation - Databricks GenAI Engineer Associate¶

Overview¶

This section covers governance and evaluation for GenAI applications, representing 20% of the exam. You need to understand RAG evaluation metrics, LLM-as-judge patterns, guardrails, and GenAI governance.

📖 Agent Evaluation - Evaluation framework 📖 AI Governance - GenAI governance

Key Topics¶

1. RAG Evaluation Dimensions¶

Key Concepts: - Retrieval quality is evaluated separately from generation quality - A good retriever with a poor generator produces grounded but poorly written answers - A poor retriever with a good generator produces fluent but potentially hallucinated answers - Evaluate both components independently to identify bottlenecks

2. Mosaic AI Agent Evaluation¶

📖 Agent Evaluation - Evaluation framework

import mlflow

# Evaluate RAG chain
eval_results = mlflow.evaluate(
    model=rag_chain,
    data=eval_dataset,
    model_type="databricks-agent"
)

# View metrics
print(eval_results.metrics)

Key Concepts: - Automated evaluation using LLM-as-judge approach - LLM judges assess quality dimensions (relevance, faithfulness, correctness) - Evaluation datasets contain questions with optional ground truth answers - Per-request quality scores for granular analysis - Human evaluation workflows for subjective quality assessment - Custom evaluation metrics for domain-specific quality requirements

3. LLM-as-Judge Pattern¶

Key Concepts: - A powerful LLM evaluates the output of the RAG system - Judge LLM receives the question, retrieved context, and generated answer - Produces scores for each evaluation dimension - More scalable than human evaluation but less accurate for nuanced judgments - Combine with human evaluation for high-stakes applications - Judge model should be at least as capable as the generation model

4. MLflow Evaluate for GenAI¶

📖 MLflow Evaluate - LLM evaluation

import mlflow

# Evaluate with specific metrics
results = mlflow.evaluate(
    model=rag_chain,
    data=eval_df,
    model_type="databricks-agent",
    evaluator_config={
        "databricks-agent": {
            "metrics": ["groundedness", "relevance", "answer_correctness"]
        }
    }
)

Available Metrics: | Metric | What It Measures | |--------|-----------------| | answer_correctness | Factual accuracy against ground truth | | faithfulness | Answer supported by retrieved context | | relevance | Answer addresses the question | | chunk_relevance | Retrieved chunks are relevant to the query |

5. Guardrails and Safety¶

Key Concepts: - Input guardrails: filter or modify user inputs before processing - Output guardrails: filter or modify model outputs before returning - Content filtering: block harmful, inappropriate, or off-topic content - PII detection and redaction in inputs and outputs - Prompt injection awareness: sanitize user inputs to prevent manipulation - Topic boundaries: constrain the model to answer only domain-relevant questions

Common Guardrail Types: | Type | Purpose | |------|---------| | Input validation | Check for malicious prompts, PII, off-topic queries | | Output filtering | Remove harmful content, PII, or hallucinated information | | Token limits | Prevent excessive cost from long prompts or responses | | Rate limiting | Protect against abuse and control costs |

6. GenAI Governance¶

📖 AI Governance - Governance patterns

Key Concepts: - Unity Catalog governs models, data, and feature assets - PII detection and handling in training data and embeddings - Data lineage tracks which documents are embedded in the vector store - Inference table logging captures all requests and responses for audit - Cost management: monitor token usage and compute costs - Data retention policies for inference logs and conversation history - Compliance: audit logging for all GenAI requests

7. Cost Management for GenAI¶

Key Concepts: - Token costs vary significantly between models (larger models cost more) - Monitor input and output token counts per request - Optimize prompt length to reduce costs (concise system prompts) - Use smaller models for simple tasks, larger models for complex tasks - Caching repeated queries can reduce token costs - Set max_tokens to prevent unexpectedly long responses

Exam Tips for This Domain¶

1. Evaluation dimensions - Know faithfulness, relevance, groundedness, correctness
2. LLM-as-judge - Understand automated evaluation and its limitations
3. Guardrails - Input/output filtering, PII handling, prompt injection
4. mlflow.evaluate - Know how to evaluate RAG chains with agent metrics
5. Governance - Unity Catalog for models, inference table logging for audit
6. Cost management - Token monitoring, model selection, prompt optimization

Documentation Links Summary¶