tests/test_analytics_contracts.py

from __future__ import annotations

from src.analytics import (
    MIN_CONFIG_N,
    config_leaderboard,
    make_decision_brief,
    overview_metrics,
    policy_at_threshold,
    policy_curve,
    retrieval_outcomes,
    risk_slices,
    top_examples,
)
from src.data import DataBundle, filter_eval, filter_retrieval_events


def test_overview_metrics_shape(bundle: DataBundle) -> None:
    metrics = overview_metrics(bundle.eval_runs, bundle.documents, bundle.chunks, bundle.retrieval_events)
    assert {"evaluations", "correct_rate", "hallucination_rate", "recall_at_10", "mrr_at_10"}.issubset(metrics)


def test_risk_slices_have_risk_score(bundle: DataBundle) -> None:
    out = risk_slices(bundle.eval_runs)
    assert not out.empty and "risk_score" in out.columns


def test_retrieval_outcomes_have_failure_modes(bundle: DataBundle) -> None:
    out = retrieval_outcomes(bundle.eval_runs)
    assert not out.empty and "failure_mode" in out.columns


def test_config_leaderboard_has_config_and_score(bundle: DataBundle) -> None:
    out = config_leaderboard(bundle.eval_runs, min_n=MIN_CONFIG_N)
    assert not out.empty and {"config", "score", "correct_rate", "hallucination_rate"}.issubset(out.columns)


def test_policy_curve_has_required_columns(bundle: DataBundle) -> None:
    out = policy_curve(bundle.eval_runs)
    assert {"threshold", "auto_approve_rate", "review_queue_size", "risk_captured_in_review"}.issubset(out.columns)


def test_policy_at_threshold_returns_decision_metrics(bundle: DataBundle) -> None:
    out = policy_at_threshold(bundle.eval_runs, 0.55, reference_df=bundle.eval_runs)
    assert {"auto_approve_rate", "review_queue_size", "auto_correct_rate", "risk_captured_in_review"}.issubset(out)


def test_decision_brief_has_action(bundle: DataBundle) -> None:
    brief = make_decision_brief(bundle.eval_runs, bundle.documents, bundle.chunks, bundle.retrieval_events)
    assert brief.recommended_action


def test_high_risk_examples_are_returned(bundle: DataBundle) -> None:
    out = top_examples(bundle.eval_runs, mode="High risk", n=20, reference_df=bundle.eval_runs)
    assert 0 < len(out) <= 20


def test_domain_filter_keeps_retrieval_alignment(bundle: DataBundle) -> None:
    filtered_eval = filter_eval(bundle.eval_runs, domains=["financial_reports"])
    filtered_retrieval = filter_retrieval_events(bundle.retrieval_events, filtered_eval)
    assert set(filtered_retrieval["example_id"].astype(str)).issubset(set(filtered_eval["example_id"].astype(str)))


def test_empty_filter_returns_empty_retrieval(bundle: DataBundle) -> None:
    filtered_eval = filter_eval(bundle.eval_runs, domains=["missing_domain"])
    filtered_retrieval = filter_retrieval_events(bundle.retrieval_events, filtered_eval)
    assert filtered_eval.empty and filtered_retrieval.empty