Want to learn evals? buildevals.com →

// EVAL CATALOG

Every eval, scorer and judge. All 47 of them.

Everything RunAgain can score a run with, from model-free string checks to calibrated LLM judges. Attach any of them online to live traffic or offline in experiments; parameters live in the docs.

heuristic scorers 12built-in llm judges 11rag & grounding judges 6retrieval ranking & search metrics 8embedding scorers 3online eval kinds 4composite & human methods 3

Heuristic scorers

Deterministic, model-free, no API key required. Each compares an output to a reference (or checks the output alone) and returns 0 to 1. docs →

exact_matchneeds expected

1 if the output equals the expected value exactly.

levenshteinneeds expected

Normalized edit-distance similarity.

numeric_diffneeds expected

Numeric closeness, where 1 means equal.

{}
json_diffneeds expected

Fraction of matching JSON leaf paths.

bleuneeds expected

Sentence BLEU: n-grams 1 to 4 plus a brevity penalty.

rouge_lneeds expected

ROUGE-L, the longest-common-subsequence F1.

chrfneeds expected

Character n-gram F-score, tolerant of word order.

token_f1needs expected

Token-overlap F1 between output and reference.

{}
valid_json

1 if the output parses as JSON.

contains

1 if the output contains a given substring.

/.*/
regex_match

1 if the output matches a given regex.

·
list_contains

Fraction of given items present in the output.

Built-in LLM judges

Ready-made judges you attach by name, no prompt to write. Each returns 0 to 1 where higher is better; for the safety judges, higher means safer. Grading uses claude-sonnet-5 by default, overridable per scorer. docs →

factualityneeds expected

Answer versus reference: subset, superset, equal or disagree.

correctnessneeds expected

Is the answer correct given the reference? Graded 0 to 1.

correctness_classifierneeds expected

Binary CORRECT or INCORRECT verdict; high precision on clear-cut errors.

completeness

Does the response cover every information need in the query? Catches partial answers.

response_quality

Is the response well organized and easy to use, regardless of correctness?

relevance

Does the response address the request?

conciseness

Is the response tight and free of filler?

toxicity

Safety: 1 means free of toxic or harmful content.

a@b.io
pii

Privacy: 1 means no PII disclosed.

prompt_injection

Security: 1 means the run resisted prompt injection.

vs
battleneeds expected

Pairwise: output versus reference in random A/B order to cancel positional bias. 1 win, 0.5 tie, 0 loss.

RAG & grounding judges

For retrieval-augmented answers: pass the retrieved passages as context and these grade grounding and retrieval quality. docs →

faithfulnessneeds context

1 means every claim is grounded in the context. Holistic, one call.

faithfulness_claimsneeds context

Decomposes the answer into atomic claims and scores the fraction supported: your hallucination rate, with per-claim verdicts.

context_relevanceneeds context

Is the retrieved context relevant to the question?

context_recallneeds context + expected

Does the context contain what the reference answer needs?

?
answer_relevancy

Is the answer responsive to the question?

source_qualityneeds context

Are the sources authoritative, relevant and current? Flags SEO spam and outdated citations.

Retrieval ranking & search metrics

Deterministic IR metrics over ranked retrieved ids, scored against binary relevance or graded qrels labels, plus model-free search checks. All accept an optional k cutoff. docs →

hit_rateneeds expected | labels

hit@k: 1 if any top-k id is relevant.

1/
mrrneeds expected | labels

Reciprocal rank of the first relevant id.

precision_at_kneeds expected | labels

Fraction of the top-k that are relevant.

recall_at_kneeds expected | labels

Fraction of all relevant refs found in the top-k.

ndcgneeds expected | labels

nDCG@k with graded gains when labels are set, else binary.

average_precisionneeds expected | labels

Per-query average precision, the term of MAP.

entity_coverage

Fraction of the query's entities (proper nouns, quoted spans, numbers) that appear in the response.

source_diversity

Unique source domains behind the response, normalized 0 to 1 by a cap argument.

Embedding scorers

Semantic similarity via your project's embedding provider (OpenAI, Cohere or Gemini). Cosine similarity rescaled into 0 to 1; embedding tokens and cost recorded per score. docs →

embedding_similarityneeds expected

Semantic closeness between output and expected, for when wording differs but meaning should match.

bertscore_relevance

Query versus response similarity: catches answers that drift off the question.

bertscore_faithfulnessneeds context

Response versus sources similarity: a fast hallucination proxy.

Online eval kinds

Applied to live production traffic, scoped to all runs, one agent, or any metadata attribute, with a coverage slider to sample a percentage of matching runs. docs →

0.9
llm_judge.<target>

A model grades the run's output or its tool calls, with the judge model of your choice.

trajectory.tool_success

Model-free score of how well the agent used its tools, with optional recovery invalidation.

structured_match

Property checks on JSON answers, plus shape-drift alerting via structured_output.changed.

baseline.*

Flags abnormal runtime, cost or span count per agent: the run baselines online eval.

Composite & human methods

For when one number is not enough: weighted multi-criterion scoring, dataset comparison, and human review that calibrates the judges. docs →

rubric

Multi-criterion scoring: each weighted dimension graded 1 to 5 in one call, with optional per-dimension minimums for pass or fail.

datasetneeds expected

Compare outputs against expected values from a versioned dataset.

annotation

Human review queues write scores next to eval scores, and the judge-alignment table measures how often each judge agrees with your reviewers.

Or bring your own judge.

Custom judge prompts, weighted rubrics and your own goal signals all land as first-class scores next to these.