Initial curated RAGcore source from c16667f0c9397e2bfc8dd0255b114fba23c8d05e

tests/evaluation/metrics/test_baseline_store.py

@@ -0,0 +1,64 @@
+"""Baseline snapshot persistence: round-trip and the committed sample fixture."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from ragcore.evaluation.metrics import AggregateMetrics, MetricsReport, load_baseline, save_baseline
+
+REPOSITORY_ROOT = Path(__file__).resolve().parents[3]
+SAMPLE_BASELINE_PATH = REPOSITORY_ROOT / "evaluation" / "baselines" / "sample-v1.baseline.json"
+
+
+def make_report() -> MetricsReport:
+    return MetricsReport(
+        dataset_id="ds",
+        dataset_version="v1",
+        overall=AggregateMetrics(
+            case_count=5,
+            mean_recall_at_5=0.8,
+            mean_recall_at_10=0.9,
+            mean_reciprocal_rank=0.7,
+            mean_ndcg_at_10=0.85,
+            zero_result_rate=0.0,
+            mean_document_diversity=2.5,
+            mean_duplicate_ratio=0.1,
+            mean_latency_ms=120.5,
+        ),
+        per_tag={
+            "exact-identifier": AggregateMetrics(
+                case_count=2,
+                mean_recall_at_5=1.0,
+                mean_recall_at_10=1.0,
+                mean_reciprocal_rank=1.0,
+                mean_ndcg_at_10=1.0,
+                zero_result_rate=0.0,
+            )
+        },
+    )
+
+
+def test_baseline_round_trips_exactly(tmp_path: Path) -> None:
+    path = tmp_path / "baseline.json"
+    report = make_report()
+
+    save_baseline(path, report)
+    recovered = load_baseline(path)
+
+    assert recovered == report
+
+
+def test_committed_sample_baseline_loads_and_validates() -> None:
+    report = load_baseline(SAMPLE_BASELINE_PATH)
+
+    assert report.dataset_id == "sample"
+    assert report.dataset_version == "v1"
+    assert report.overall.case_count == 3
+    assert "exact-identifier" in report.per_tag
+
+
+def test_reloading_the_committed_baseline_is_deterministic() -> None:
+    first = load_baseline(SAMPLE_BASELINE_PATH)
+    second = load_baseline(SAMPLE_BASELINE_PATH)
+
+    assert first == second

tests/evaluation/metrics/test_calculator.py

@@ -0,0 +1,188 @@
+"""compute_case_metrics / aggregate_metrics against a small evaluation run."""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from uuid import uuid4
+
+import pytest
+
+from ragcore.evaluation import (
+    AnswerabilityStatus,
+    CaseResult,
+    EvaluationCase,
+    EvaluationDataset,
+    EvaluationRun,
+    RetrievalRunProfile,
+)
+from ragcore.evaluation.metrics import MetricsError, aggregate_metrics, compute_case_metrics
+
+
+def make_profile() -> RetrievalRunProfile:
+    return RetrievalRunProfile(
+        embedding_profile_digest="a" * 64,
+        sparse_profile_version="v1",
+        index_alias="rag_dense_hybrid_active",
+        index_collection_version="v1",
+    )
+
+
+def test_case_with_no_relevant_chunks_is_excluded_from_scoring() -> None:
+    case = EvaluationCase(
+        id="no-evidence",
+        question="q",
+        language="en",
+        expected_answerability=AnswerabilityStatus.NOT_ANSWERABLE,
+    )
+    result = CaseResult(
+        case_id="no-evidence",
+        retrieved_chunk_ids=(),
+        matched_expected_ids=(),
+        missing_expected_ids=(),
+        unexpected_ids=(),
+    )
+    assert compute_case_metrics(case, result) is None
+
+
+def test_case_metrics_document_diversity_and_duplicate_ratio() -> None:
+    chunk_a, chunk_b, chunk_c = uuid4(), uuid4(), uuid4()
+    document_x, document_y = uuid4(), uuid4()
+    case = EvaluationCase(
+        id="case-1",
+        question="q",
+        language="en",
+        expected_answerability=AnswerabilityStatus.ANSWERABLE,
+        expected_chunk_ids=(chunk_a,),
+        expected_document_ids=(document_x,),
+        required_facts=("fact",),
+    )
+    result = CaseResult(
+        case_id="case-1",
+        retrieved_chunk_ids=(chunk_a, chunk_b, chunk_c),
+        matched_expected_ids=(chunk_a,),
+        missing_expected_ids=(),
+        unexpected_ids=(chunk_b, chunk_c),
+    )
+    document_ids = {chunk_a: document_x, chunk_b: document_x, chunk_c: document_y}
+
+    metrics = compute_case_metrics(case, result, document_ids=document_ids)
+
+    assert metrics is not None
+    assert metrics.document_diversity == 2  # document_x and document_y
+    assert metrics.duplicate_ratio == pytest.approx(1 / 3)  # 1 duplicate out of 3
+    assert metrics.recall_at_5 == 1.0
+    assert metrics.reciprocal_rank == 1.0
+
+
+def test_aggregate_metrics_overall_and_per_tag() -> None:
+    matched_chunk = uuid4()
+    document_id = uuid4()
+    answerable_case = EvaluationCase(
+        id="answerable-1",
+        question="q1",
+        language="en",
+        expected_answerability=AnswerabilityStatus.ANSWERABLE,
+        expected_chunk_ids=(matched_chunk,),
+        expected_document_ids=(document_id,),
+        required_facts=("fact",),
+        tags=("exact-identifier",),
+    )
+    no_evidence_case = EvaluationCase(
+        id="no-evidence-1",
+        question="q2",
+        language="en",
+        expected_answerability=AnswerabilityStatus.NOT_ANSWERABLE,
+        tags=("out-of-scope",),
+    )
+    dataset = EvaluationDataset(id="ds", version="v1", cases=(answerable_case, no_evidence_case))
+    run = EvaluationRun(
+        id=uuid4(),
+        dataset_id="ds",
+        dataset_version="v1",
+        run_profile=make_profile(),
+        started_at=datetime(2026, 8, 2, tzinfo=UTC),
+        case_results=(
+            CaseResult(
+                case_id="answerable-1",
+                retrieved_chunk_ids=(matched_chunk,),
+                matched_expected_ids=(matched_chunk,),
+                missing_expected_ids=(),
+                unexpected_ids=(),
+            ),
+            CaseResult(
+                case_id="no-evidence-1",
+                retrieved_chunk_ids=(),
+                matched_expected_ids=(),
+                missing_expected_ids=(),
+                unexpected_ids=(),
+            ),
+        ),
+    )
+
+    report = aggregate_metrics(dataset, run)
+
+    assert report.overall.case_count == 1  # only the answerable case is scored
+    assert report.overall.mean_recall_at_10 == 1.0
+    assert report.overall.zero_result_rate == pytest.approx(0.5)  # 1 of 2 cases retrieved nothing
+    assert "exact-identifier" in report.per_tag
+    assert report.per_tag["exact-identifier"].mean_recall_at_10 == 1.0
+    assert "out-of-scope" not in report.per_tag  # no scored case carries this tag
+
+
+def test_aggregate_metrics_rejects_a_run_for_a_different_dataset() -> None:
+    case = EvaluationCase(
+        id="c",
+        question="q",
+        language="en",
+        expected_answerability=AnswerabilityStatus.ANSWERABLE,
+        expected_chunk_ids=(uuid4(),),
+        expected_document_ids=(uuid4(),),
+        required_facts=("fact",),
+    )
+    dataset = EvaluationDataset(id="ds", version="v1", cases=(case,))
+    run = EvaluationRun(
+        id=uuid4(),
+        dataset_id="other",
+        dataset_version="v1",
+        run_profile=make_profile(),
+        started_at=datetime(2026, 8, 2, tzinfo=UTC),
+        case_results=(
+            CaseResult(
+                case_id="c",
+                retrieved_chunk_ids=(),
+                matched_expected_ids=(),
+                missing_expected_ids=(),
+                unexpected_ids=(),
+            ),
+        ),
+    )
+    with pytest.raises(MetricsError, match="does not belong"):
+        aggregate_metrics(dataset, run)
+
+
+def test_aggregate_metrics_requires_at_least_one_scoreable_case() -> None:
+    case = EvaluationCase(
+        id="c",
+        question="q",
+        language="en",
+        expected_answerability=AnswerabilityStatus.NOT_ANSWERABLE,
+    )
+    dataset = EvaluationDataset(id="ds", version="v1", cases=(case,))
+    run = EvaluationRun(
+        id=uuid4(),
+        dataset_id="ds",
+        dataset_version="v1",
+        run_profile=make_profile(),
+        started_at=datetime(2026, 8, 2, tzinfo=UTC),
+        case_results=(
+            CaseResult(
+                case_id="c",
+                retrieved_chunk_ids=(),
+                matched_expected_ids=(),
+                missing_expected_ids=(),
+                unexpected_ids=(),
+            ),
+        ),
+    )
+    with pytest.raises(MetricsError, match="relevant chunks to score"):
+        aggregate_metrics(dataset, run)

tests/evaluation/metrics/test_formulas.py

@@ -0,0 +1,80 @@
+"""Metric formulas verified against hand-computed golden values."""
+
+from __future__ import annotations
+
+import math
+from uuid import UUID
+
+import pytest
+
+from ragcore.evaluation.metrics import ndcg_at_k, recall_at_k, reciprocal_rank
+
+D1, D2, D3, D4, D5 = (
+    UUID(f"00000000-0000-4000-8000-00000000000{i}") for i in range(1, 6)
+)
+
+
+def test_recall_at_k_golden_values() -> None:
+    retrieved = (D1, D2, D3, D4, D5)
+    relevant = {D2, D4, D5}
+
+    # Top 2: {D1, D2} intersect relevant = {D2} -> 1/3
+    assert recall_at_k(retrieved, relevant, 2) == pytest.approx(1 / 3)
+    # Top 4: {D1, D2, D3, D4} intersect relevant = {D2, D4} -> 2/3
+    assert recall_at_k(retrieved, relevant, 4) == pytest.approx(2 / 3)
+    # Top 5: all relevant found -> 1.0
+    assert recall_at_k(retrieved, relevant, 5) == pytest.approx(1.0)
+
+
+def test_recall_at_k_with_no_hits_is_zero() -> None:
+    assert recall_at_k((D1, D3), {D2, D4}, 2) == 0.0
+
+
+def test_recall_at_k_requires_nonempty_relevant() -> None:
+    with pytest.raises(ValueError, match="relevant"):
+        recall_at_k((D1,), set(), 5)
+
+
+def test_reciprocal_rank_golden_values() -> None:
+    # First relevant item is at rank 3 -> 1/3.
+    assert reciprocal_rank((D1, D3, D2, D4), {D2, D4}) == pytest.approx(1 / 3)
+    # First relevant item is at rank 1 -> 1.0.
+    assert reciprocal_rank((D2, D1, D3), {D2}) == pytest.approx(1.0)
+    # No relevant item retrieved -> 0.0.
+    assert reciprocal_rank((D1, D3), {D2}) == 0.0
+
+
+def test_ndcg_at_k_golden_value_hand_computed() -> None:
+    # retrieved = [D1(rel), D2(rel), D3(non-rel), D4(rel)], relevant = {D1, D2, D4}, k=4.
+    # DCG@4  = 1/log2(2) + 1/log2(3) + 0 + 1/log2(5)
+    # IDCG@4 = 1/log2(2) + 1/log2(3) + 1/log2(4)   (3 relevant items, ideal order)
+    retrieved = (D1, D2, D3, D4)
+    relevant = {D1, D2, D4}
+
+    dcg = 1 / math.log2(2) + 1 / math.log2(3) + 1 / math.log2(5)
+    idcg = 1 / math.log2(2) + 1 / math.log2(3) + 1 / math.log2(4)
+    expected = dcg / idcg
+
+    assert ndcg_at_k(retrieved, relevant, 4) == pytest.approx(expected)
+    # Hand-verified numeric value to guard against a formula-shape regression
+    # (e.g. an off-by-one rank or a wrong log base) that the symbolic
+    # recomputation above would not catch.
+    assert ndcg_at_k(retrieved, relevant, 4) == pytest.approx(0.967468, abs=1e-5)
+
+
+def test_ndcg_at_k_perfect_ranking_is_one() -> None:
+    retrieved = (D1, D2, D3)
+    relevant = {D1, D2, D3}
+    assert ndcg_at_k(retrieved, relevant, 3) == pytest.approx(1.0)
+
+
+def test_ndcg_at_k_no_relevant_hits_is_zero() -> None:
+    assert ndcg_at_k((D1, D3), {D2, D4}, 2) == 0.0
+
+
+def test_formulas_reject_oversized_input() -> None:
+    from ragcore.evaluation.metrics import MAX_RANKED_ITEMS
+
+    oversized = tuple(D1 for _ in range(MAX_RANKED_ITEMS + 1))
+    with pytest.raises(ValueError, match="bounded limit"):
+        recall_at_k(oversized, {D1}, 5)

tests/evaluation/metrics/test_gate.py

@@ -0,0 +1,147 @@
+"""evaluate_gate: baseline-vs-candidate comparison and per-tag regression detection."""
+
+from __future__ import annotations
+
+import pytest
+
+from ragcore.evaluation.metrics import (
+    AggregateMetrics,
+    MetricsError,
+    MetricsReport,
+    RegressionMetric,
+    RegressionThresholds,
+    evaluate_gate,
+)
+
+
+def make_aggregate(**changes: object) -> AggregateMetrics:
+    value: dict[str, object] = {
+        "case_count": 10,
+        "mean_recall_at_5": 0.8,
+        "mean_recall_at_10": 0.9,
+        "mean_reciprocal_rank": 0.7,
+        "mean_ndcg_at_10": 0.85,
+        "zero_result_rate": 0.05,
+    }
+    value.update(changes)
+    return AggregateMetrics(**value)
+
+
+def make_report(
+    overall: AggregateMetrics, per_tag: dict[str, AggregateMetrics] | None = None
+) -> MetricsReport:
+    return MetricsReport(
+        dataset_id="ds", dataset_version="v1", overall=overall, per_tag=per_tag or {}
+    )
+
+
+def default_thresholds(**changes: object) -> RegressionThresholds:
+    value: dict[str, object] = {"version": "gate-v1"}
+    value.update(changes)
+    return RegressionThresholds(**value)
+
+
+def test_identical_metrics_pass_the_gate() -> None:
+    metrics = make_aggregate()
+    result = evaluate_gate(make_report(metrics), make_report(metrics), default_thresholds())
+
+    assert result.passed
+    assert result.regressions == ()
+
+
+def test_improved_metrics_pass_the_gate() -> None:
+    baseline = make_aggregate(mean_recall_at_10=0.8)
+    candidate = make_aggregate(mean_recall_at_10=0.95)
+    result = evaluate_gate(make_report(baseline), make_report(candidate), default_thresholds())
+
+    assert result.passed
+
+
+def test_recall_drop_beyond_threshold_fails_the_gate() -> None:
+    baseline = make_aggregate(mean_recall_at_10=0.9)
+    candidate = make_aggregate(mean_recall_at_10=0.80)  # 0.10 drop > default 0.05
+    result = evaluate_gate(make_report(baseline), make_report(candidate), default_thresholds())
+
+    assert not result.passed
+    assert any(f.metric is RegressionMetric.RECALL_AT_10 for f in result.regressions)
+
+
+def test_recall_drop_within_threshold_passes() -> None:
+    baseline = make_aggregate(mean_recall_at_10=0.90)
+    candidate = make_aggregate(mean_recall_at_10=0.87)  # 0.03 drop <= default 0.05
+    result = evaluate_gate(make_report(baseline), make_report(candidate), default_thresholds())
+
+    assert result.passed
+
+
+def test_zero_result_rate_rise_fails_the_gate() -> None:
+    baseline = make_aggregate(zero_result_rate=0.02)
+    candidate = make_aggregate(zero_result_rate=0.10)  # 0.08 rise > default 0.02
+    result = evaluate_gate(make_report(baseline), make_report(candidate), default_thresholds())
+
+    assert not result.passed
+    assert any(f.metric is RegressionMetric.ZERO_RESULT_RATE for f in result.regressions)
+
+
+def test_per_tag_regression_is_highlighted_even_when_overall_passes() -> None:
+    baseline = make_report(
+        make_aggregate(),
+        per_tag={
+            "exact-identifier": make_aggregate(mean_recall_at_10=0.95),
+            "conceptual": make_aggregate(mean_recall_at_10=0.85),
+        },
+    )
+    candidate = make_report(
+        make_aggregate(),  # overall unchanged
+        per_tag={
+            "exact-identifier": make_aggregate(mean_recall_at_10=0.70),  # regressed
+            "conceptual": make_aggregate(mean_recall_at_10=0.86),  # slightly improved
+        },
+    )
+
+    result = evaluate_gate(baseline, candidate, default_thresholds())
+
+    assert not result.passed
+    tag_regressions = {finding.tag for finding in result.regressions}
+    assert "exact-identifier" in tag_regressions
+    assert "conceptual" not in tag_regressions
+
+
+def test_tags_only_present_in_one_report_are_not_compared() -> None:
+    baseline = make_report(make_aggregate(), per_tag={"old-tag": make_aggregate()})
+    candidate = make_report(
+        make_aggregate(), per_tag={"new-tag": make_aggregate(mean_recall_at_10=0.0)}
+    )
+
+    result = evaluate_gate(baseline, candidate, default_thresholds())
+
+    assert result.passed  # only overall is compared; mismatched tags are skipped
+
+
+def test_gate_rejects_mismatched_datasets() -> None:
+    baseline = MetricsReport(dataset_id="a", dataset_version="v1", overall=make_aggregate())
+    candidate = MetricsReport(dataset_id="b", dataset_version="v1", overall=make_aggregate())
+
+    with pytest.raises(MetricsError, match="same dataset"):
+        evaluate_gate(baseline, candidate, default_thresholds())
+
+
+def test_gate_result_invariants_are_enforced() -> None:
+    from ragcore.evaluation.metrics import GateResult, RegressionFinding
+
+    with pytest.raises(MetricsError, match="must not carry regressions"):
+        GateResult(
+            thresholds_version="v1",
+            passed=True,
+            regressions=(
+                RegressionFinding(
+                    metric=RegressionMetric.RECALL_AT_5,
+                    tag=None,
+                    baseline_value=0.9,
+                    candidate_value=0.5,
+                    delta=-0.4,
+                ),
+            ),
+        )
+    with pytest.raises(MetricsError, match="at least one regression"):
+        GateResult(thresholds_version="v1", passed=False, regressions=())