Initial curated RAGcore source from c16667f0c9397e2bfc8dd0255b114fba23c8d05e

tests/evaluation/framework/test_case_schema.py

@@ -0,0 +1,119 @@
+"""Versioned case/dataset schema validation."""
+
+from __future__ import annotations
+
+from uuid import uuid4
+
+import pytest
+
+from ragcore.evaluation import (
+    AnswerabilityStatus,
+    EvaluationCase,
+    EvaluationDataset,
+    EvaluationDatasetError,
+    RetrievalRunProfile,
+)
+
+
+def make_case(**changes: object) -> EvaluationCase:
+    value: dict[str, object] = {
+        "id": "case-1",
+        "question": "What is the default timeout?",
+        "language": "en",
+        "expected_answerability": AnswerabilityStatus.ANSWERABLE,
+        "expected_document_ids": (uuid4(),),
+        "expected_chunk_ids": (uuid4(),),
+        "required_facts": ("the default timeout is 30 seconds",),
+    }
+    value.update(changes)
+    return EvaluationCase(**value)
+
+
+def test_valid_answerable_case_is_accepted() -> None:
+    case = make_case()
+    assert case.expected_answerability is AnswerabilityStatus.ANSWERABLE
+
+
+def test_answerable_case_requires_expected_documents() -> None:
+    with pytest.raises(EvaluationDatasetError, match="expect at least one document"):
+        make_case(expected_document_ids=())
+
+
+def test_answerable_case_requires_required_facts() -> None:
+    with pytest.raises(EvaluationDatasetError, match="required facts"):
+        make_case(required_facts=())
+
+
+def test_not_answerable_case_may_have_no_expected_evidence() -> None:
+    case = make_case(
+        expected_answerability=AnswerabilityStatus.NOT_ANSWERABLE,
+        expected_document_ids=(),
+        expected_chunk_ids=(),
+        required_facts=(),
+    )
+    assert case.expected_document_ids == ()
+
+
+def test_conflicting_evidence_case_requires_two_distinct_documents() -> None:
+    with pytest.raises(EvaluationDatasetError, match="two distinct documents"):
+        make_case(
+            expected_answerability=AnswerabilityStatus.CONFLICTING_EVIDENCE,
+            expected_document_ids=(uuid4(),),
+            required_facts=(),
+        )
+
+    case = make_case(
+        expected_answerability=AnswerabilityStatus.CONFLICTING_EVIDENCE,
+        expected_document_ids=(uuid4(), uuid4()),
+        required_facts=(),
+    )
+    assert case.expected_answerability is AnswerabilityStatus.CONFLICTING_EVIDENCE
+
+
+def test_case_id_and_language_are_validated() -> None:
+    with pytest.raises(EvaluationDatasetError, match="identifier"):
+        make_case(id="Not A Safe Id")
+    with pytest.raises(EvaluationDatasetError, match="language"):
+        make_case(language="english")
+
+
+def test_case_forbids_unbounded_collections() -> None:
+    with pytest.raises(EvaluationDatasetError, match="exceeds the bounded limit"):
+        make_case(required_facts=tuple(f"fact {i}" for i in range(33)))
+
+
+def test_dataset_requires_at_least_one_case() -> None:
+    with pytest.raises(EvaluationDatasetError, match="between 1 and 500"):
+        EvaluationDataset(id="empty", version="v1", cases=())
+
+
+def test_dataset_rejects_duplicate_case_ids() -> None:
+    with pytest.raises(EvaluationDatasetError, match="unique"):
+        EvaluationDataset(id="dup", version="v1", cases=(make_case(), make_case()))
+
+
+def test_dataset_id_and_version_are_validated() -> None:
+    with pytest.raises(EvaluationDatasetError, match="identifier"):
+        EvaluationDataset(id="Not Safe", version="v1", cases=(make_case(),))
+    with pytest.raises(EvaluationDatasetError, match="identifier"):
+        EvaluationDataset(id="dataset", version="Not Safe", cases=(make_case(),))
+
+
+def test_run_profile_fields_are_bounded_and_non_empty() -> None:
+    with pytest.raises(EvaluationDatasetError, match="embedding_profile_digest"):
+        RetrievalRunProfile(
+            embedding_profile_digest="",
+            sparse_profile_version="v1",
+            index_alias="rag_dense_x_active",
+            index_collection_version="v1",
+        )
+
+
+def test_run_profile_allows_no_reranker() -> None:
+    profile = RetrievalRunProfile(
+        embedding_profile_digest="a" * 64,
+        sparse_profile_version="v1",
+        index_alias="rag_dense_x_active",
+        index_collection_version="v1",
+    )
+    assert profile.reranker_profile_digest is None

tests/evaluation/framework/test_loader.py

@@ -0,0 +1,79 @@
+"""Loading and validating dataset JSON files against the versioned schema."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from ragcore.evaluation import AnswerabilityStatus, EvaluationDatasetError, load_dataset
+
+REPOSITORY_ROOT = Path(__file__).resolve().parents[3]
+SAMPLE_DATASET_PATH = REPOSITORY_ROOT / "evaluation" / "datasets" / "sample-v1.json"
+
+
+def test_sample_fixture_dataset_loads_and_validates() -> None:
+    dataset = load_dataset(SAMPLE_DATASET_PATH)
+
+    assert dataset.id == "sample"
+    assert dataset.version == "v1"
+    assert len(dataset.cases) == 4
+    case_by_id = {case.id: case for case in dataset.cases}
+    assert case_by_id["exact-error-code-lookup"].expected_answerability is (
+        AnswerabilityStatus.ANSWERABLE
+    )
+    assert case_by_id["conflicting-version-docs"].expected_answerability is (
+        AnswerabilityStatus.CONFLICTING_EVIDENCE
+    )
+    assert case_by_id["no-relevant-evidence"].expected_document_ids == ()
+
+
+def test_loader_rejects_missing_file() -> None:
+    with pytest.raises(EvaluationDatasetError, match="could not be read"):
+        load_dataset(REPOSITORY_ROOT / "evaluation" / "datasets" / "does-not-exist.json")
+
+
+def test_loader_rejects_invalid_json(tmp_path: Path) -> None:
+    path = tmp_path / "broken.json"
+    path.write_text("{not valid json", encoding="utf-8")
+    with pytest.raises(EvaluationDatasetError, match="valid UTF-8 JSON"):
+        load_dataset(path)
+
+
+def test_loader_rejects_non_object_payload(tmp_path: Path) -> None:
+    path = tmp_path / "list.json"
+    path.write_text(json.dumps(["not", "an", "object"]), encoding="utf-8")
+    with pytest.raises(EvaluationDatasetError, match="JSON object"):
+        load_dataset(path)
+
+
+def test_loader_rejects_malformed_uuid(tmp_path: Path) -> None:
+    path = tmp_path / "bad-uuid.json"
+    path.write_text(
+        json.dumps(
+            {
+                "id": "bad",
+                "version": "v1",
+                "cases": [
+                    {
+                        "id": "case-1",
+                        "question": "q",
+                        "language": "en",
+                        "expected_answerability": "not_answerable",
+                        "expected_document_ids": ["not-a-uuid"],
+                    }
+                ],
+            }
+        ),
+        encoding="utf-8",
+    )
+    with pytest.raises(EvaluationDatasetError, match="UUID strings"):
+        load_dataset(path)
+
+
+def test_loader_reloading_the_same_file_is_deterministic() -> None:
+    first = load_dataset(SAMPLE_DATASET_PATH)
+    second = load_dataset(SAMPLE_DATASET_PATH)
+
+    assert first == second

tests/evaluation/framework/test_runner.py

@@ -0,0 +1,119 @@
+"""EvaluationRunner: offline fixture-corpus runs with reproducible results."""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from pathlib import Path
+from uuid import uuid4
+
+import pytest
+
+from ragcore.evaluation import (
+    AnswerabilityStatus,
+    EvaluationCase,
+    EvaluationDataset,
+    EvaluationDatasetError,
+    EvaluationRunner,
+    FixtureCorpusRetriever,
+    RetrievalRunProfile,
+)
+
+REPOSITORY_ROOT = Path(__file__).resolve().parents[3]
+
+
+def make_dataset() -> tuple[EvaluationDataset, dict]:
+    matched_chunk, missing_chunk, unexpected_chunk = uuid4(), uuid4(), uuid4()
+    case = EvaluationCase(
+        id="case-1",
+        question="What is the timeout?",
+        language="en",
+        expected_answerability=AnswerabilityStatus.ANSWERABLE,
+        expected_document_ids=(uuid4(),),
+        expected_chunk_ids=(matched_chunk, missing_chunk),
+        required_facts=("the timeout is bounded",),
+    )
+    dataset = EvaluationDataset(id="offline", version="v1", cases=(case,))
+    corpus = {"case-1": (matched_chunk, unexpected_chunk)}
+    return dataset, {
+        "corpus": corpus,
+        "matched": matched_chunk,
+        "missing": missing_chunk,
+        "unexpected": unexpected_chunk,
+    }
+
+
+def make_profile() -> RetrievalRunProfile:
+    return RetrievalRunProfile(
+        embedding_profile_digest="a" * 64,
+        sparse_profile_version="lexical-v1",
+        index_alias="rag_dense_hybrid_active",
+        index_collection_version="v1",
+    )
+
+
+async def test_runner_works_offline_with_a_fixture_corpus() -> None:
+    dataset, ids = make_dataset()
+    retriever = FixtureCorpusRetriever(ids["corpus"])
+    runner = EvaluationRunner(retriever, clock=lambda: datetime(2026, 8, 2, tzinfo=UTC))
+
+    run = await runner.run(dataset, run_profile=make_profile())
+
+    assert run.dataset_id == "offline"
+    assert run.dataset_version == "v1"
+    assert len(run.case_results) == 1
+    result = run.case_results[0]
+    assert result.matched_expected_ids == (ids["matched"],)
+    assert result.missing_expected_ids == (ids["missing"],)
+    assert result.unexpected_ids == (ids["unexpected"],)
+
+
+async def test_run_records_profile_model_and_index_revisions() -> None:
+    dataset, ids = make_dataset()
+    runner = EvaluationRunner(FixtureCorpusRetriever(ids["corpus"]))
+    profile = RetrievalRunProfile(
+        embedding_profile_digest="b" * 64,
+        sparse_profile_version="lexical-v2",
+        index_alias="rag_dense_hybrid_active",
+        index_collection_version="v3",
+        reranker_profile_digest="c" * 64,
+    )
+
+    run = await runner.run(dataset, run_profile=profile)
+
+    assert run.run_profile.embedding_profile_digest == "b" * 64
+    assert run.run_profile.sparse_profile_version == "lexical-v2"
+    assert run.run_profile.index_collection_version == "v3"
+    assert run.run_profile.reranker_profile_digest == "c" * 64
+
+
+async def test_results_are_reproducible_given_identical_inputs() -> None:
+    dataset, ids = make_dataset()
+    run_id = uuid4()
+    fixed_time = datetime(2026, 8, 2, 9, tzinfo=UTC)
+    runner_a = EvaluationRunner(FixtureCorpusRetriever(ids["corpus"]), clock=lambda: fixed_time)
+    runner_b = EvaluationRunner(FixtureCorpusRetriever(ids["corpus"]), clock=lambda: fixed_time)
+
+    run_a = await runner_a.run(dataset, run_profile=make_profile(), run_id=run_id)
+    run_b = await runner_b.run(dataset, run_profile=make_profile(), run_id=run_id)
+
+    assert run_a == run_b
+
+
+async def test_runner_rejects_a_retriever_exceeding_the_bounded_limit() -> None:
+    dataset, _ids = make_dataset()
+    oversized = {"case-1": tuple(uuid4() for _ in range(201))}
+    # FixtureCorpusRetriever itself would reject this at construction time; a
+    # raw retriever proves the runner independently enforces the limit too,
+    # regardless of which Reranker-style implementation is plugged in.
+    runner = EvaluationRunner(_RawRetriever(oversized))
+
+    with pytest.raises(EvaluationDatasetError, match="bounded limit"):
+        await runner.run(dataset, run_profile=make_profile())
+
+
+class _RawRetriever:
+    def __init__(self, corpus: dict) -> None:
+        self._corpus = corpus
+
+    async def retrieve(self, case: EvaluationCase) -> tuple:
+        return self._corpus.get(case.id, ())

tests/evaluation/framework/test_store.py

@@ -0,0 +1,75 @@
+"""JsonEvaluationRunStore: atomic persistence and exact round-trip."""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from pathlib import Path
+from uuid import uuid4
+
+import pytest
+
+from ragcore.evaluation import (
+    CaseResult,
+    EvaluationDatasetError,
+    EvaluationRun,
+    JsonEvaluationRunStore,
+    RetrievalRunProfile,
+)
+
+
+def make_run(run_id=None) -> EvaluationRun:
+    return EvaluationRun(
+        id=run_id or uuid4(),
+        dataset_id="offline",
+        dataset_version="v1",
+        run_profile=RetrievalRunProfile(
+            embedding_profile_digest="a" * 64,
+            sparse_profile_version="lexical-v1",
+            index_alias="rag_dense_hybrid_active",
+            index_collection_version="v1",
+        ),
+        started_at=datetime(2026, 8, 2, tzinfo=UTC),
+        case_results=(
+            CaseResult(
+                case_id="case-1",
+                retrieved_chunk_ids=(uuid4(),),
+                matched_expected_ids=(),
+                missing_expected_ids=(uuid4(),),
+                unexpected_ids=(),
+            ),
+        ),
+    )
+
+
+def test_stored_run_round_trips_exactly(tmp_path: Path) -> None:
+    store = JsonEvaluationRunStore(tmp_path / "runs.json")
+    run = make_run()
+
+    store.add(run)
+
+    recovered_store = JsonEvaluationRunStore(tmp_path / "runs.json")
+    assert recovered_store.get(run.id) == run
+
+
+def test_missing_run_returns_none(tmp_path: Path) -> None:
+    store = JsonEvaluationRunStore(tmp_path / "runs.json")
+    assert store.get(uuid4()) is None
+
+
+def test_duplicate_run_id_is_rejected(tmp_path: Path) -> None:
+    store = JsonEvaluationRunStore(tmp_path / "runs.json")
+    run = make_run()
+    store.add(run)
+
+    with pytest.raises(EvaluationDatasetError, match="already stored"):
+        store.add(run)
+
+
+def test_all_returns_every_stored_run(tmp_path: Path) -> None:
+    store = JsonEvaluationRunStore(tmp_path / "runs.json")
+    first, second = make_run(), make_run()
+    store.add(first)
+    store.add(second)
+
+    stored_ids = {run.id for run in store.all()}
+    assert stored_ids == {first.id, second.id}