Struct ConversationScorer 

pub struct ConversationScorer { /* private fields */ }

Available on crate feature eval only.

Scores multi-turn conversations on quality metrics.

Uses a StructuredJudge for context retention, goal completion, and coherence metrics. For topic drift, uses an [EmbeddingScorer] if available, otherwise falls back to the structured judge.

Implementations

impl ConversationScorer

pub fn new(judge: StructuredJudge) -> ConversationScorer

Create a new conversation scorer with default configuration.

Uses the structured judge for all metrics including topic drift.

pub fn with_config( judge: StructuredJudge, config: ConversationScorerConfig, ) -> ConversationScorer

Available on non-crate feature embedding only.

Create a conversation scorer with custom configuration (no embedding).

pub fn config(&self) -> &ConversationScorerConfig

Returns the current configuration.

pub async fn score( &self, conversation: &[Content], goal: &str, ) -> Result<ConversationMetrics, EvalError>

Score a multi-turn conversation.

Evaluates the conversation on four metrics:

• Context retention: Does the agent reference prior turn information?
• Goal completion: Does the agent achieve the stated objective?
• Coherence: Are consecutive responses logically consistent?
• Topic drift: Does the conversation stay on topic?

All scores are clamped to [0.0, 1.0].

Errors

Returns an error if the judge LLM calls fail or the conversation is empty.