[FR] Additional built-in LLM judges for safety, coherence, agent planning, ...

Willingness to contribute

I cannot contribute this myself, and am requesting help from other contributors

Description

Expand MLflow's built-in judge library.

These judges should be production-ready, well-tested, and work out-of-the-box with minimal configuration.

Proposed Solution

Add the following built-in judges to mlflow.genai.scorers:

1. Conversational Safety (multi-turn)

mlflow.genai.scorers.ConversationSafety()

• Goal: Evaluate safety across the entire conversation
• Outcome: Binary pass/fail with identification of safety concerns
• Rationale: Critical for production deployment of conversational agents

2. Conversational Tool Call Efficiency (multi-turn)

mlflow.genai.scorers.ConversationalToolCallEfficiency()

• Goal: Assess tool usage efficiency across the full conversation session
• Outcome: Binary pass/fail on session-level tool call optimization
• Rationale: Extends single-turn efficiency to full conversation context

3. Conversational Role Adherence (multi-turn)

mlflow.genai.scorers.ConversationalRoleAdherence()

• Goal: Verify agent maintains its assigned role throughout the conversation
• Outcome: Binary pass/fail on role consistency
• Rationale: Ensures agents stay within defined boundaries and personas

5. Conversational Coherence (multi-turn)

mlflow.genai.scorers.ConversationalCoherence()

• Goal: Assess logical flow and consistency across the entire conversation
• Outcome: Binary pass/fail on session-level coherence
• Rationale: Multi-turn extension of single-turn coherence evaluation

6. Agent Plan Quality (multi-turn)

mlflow.genai.scorers.AgentPlanQuality()

• Goal: Evaluate the quality of agent's action planning and reasoning
• Outcome: Binary pass/fail with assessment of planning effectiveness
• Rationale: Important for agentic systems that decompose tasks into steps

Implementation Considerations

1. Consistent API: All judges should follow the same calling convention as existing built-in judges
2. Documentation: Comprehensive examples showing when to use each judge

Related

• Extends #19056 (additional built-in judges)
• Builds on #19052 (make_judge for conversations)
• Works with #19055 (offline evaluation)
• Integrates with #19058 (judge builder UI)