MetricResult — standardized metric output

Every metric class in TrustifAI returns a MetricResult. The structure is identical across all four offline metrics and the confidence metric, so downstream code can process results uniformly regardless of which metric produced them. The .to_dict() method serializes the result to a plain dict suitable for JSON responses, dataframe rows, or logging.

Fields

float

required

Normalized trust score in the range [0.0, 1.0]. Higher values indicate stronger trustworthiness. The meaning of the scale is metric-specific — see the offline metrics reference for per-metric interpretation.

str

required

Human-readable classification of the score, produced by ThresholdEvaluator using thresholds from your config. Common label values include:

"Strong Grounding" / "Partial Grounding" / "Likely Hallucinated Answer" — evidence coverage and semantic drift
"Stable Consistency" / "Fragile Consistency" / "Unreliable" — epistemic consistency
"High Trust" / "Moderate Trust" / "Low Trust" — source diversity
"High Confidence" / "Medium Confidence" / "Low Confidence" — confidence

dict

required

Metric-specific breakdown of how the score was derived. Keys vary by metric. All details dicts include an "explanation" string when grading succeeded.

Show properties

str

One-sentence summary of why the score is what it is, generated by ThresholdEvaluator.

int

(EvidenceCoverageMetric only) Number of sentences in the answer that were evaluated.

int

(EvidenceCoverageMetric only) Number of sentences judged to be supported by the retrieved documents.

list

(EvidenceCoverageMetric only) The exact answer sentences that lacked document support.

str

(SemanticDriftMetric only) The document sentence with the highest cosine similarity to the answer.

list

(EpistemicConsistencyMetric only) The k stochastic re-generations used to compute the consistency score.

int

(SourceDiversityMetric only) Count of distinct source identifiers found across the documents.

dict | None

default:"None"

Present when the metric incurred external API costs. Omitted from .to_dict() output when None.

Show properties

float

Estimated cost in US dollars for LLM and embedding calls made during this metric calculation.

`.to_dict()` method

Serializes the result to a plain Python dict. The score is rounded to 2 decimal places. execution_metadata is included only if it is not None.

result.to_dict() -> dict

Example output

The following shows a MetricResult returned by EvidenceCoverageMetric and its .to_dict() representation:

result = MetricResult(
    score=0.857142857,
    label="Strong Grounding",
    details={
        "explanation": "Fully supported by source documents.",
        "total_sentences": 7,
        "supported_sentences": 6,
        "unsupported_sentences": ["The tower was built overnight."],
        "failed_checks": 0,
        "failed_reason": None,
    },
    execution_metadata={"total_cost_usd": 0.00031},
)

result.to_dict()
# {
#     "score": 0.86,
#     "label": "Strong Grounding",
#     "details": {
#         "explanation": "Fully supported by source documents.",
#         "total_sentences": 7,
#         "supported_sentences": 6,
#         "unsupported_sentences": ["The tower was built overnight."],
#         "failed_checks": 0,
#         "failed_reason": None,
#     },
#     "execution_metadata": {"total_cost_usd": 0.00031},
# }

Use .to_dict() when building JSON API responses or writing results to a database. The rounded score and flat structure make it safe to serialize with json.dumps() without further processing.

Core API

Data Structures

Metrics

MetricResult — standardized metric output

Fields

`.to_dict()` method

Example output

​Fields

​.to_dict() method

​Example output

Fields

`.to_dict()` method

Example output