new test

test/test_rag_metrics_semantic.py

@@ -144,3 +144,111 @@ def test_cesar_goal_matches_rozoznavat_variant() -> None:
         "recognize unknown named entities",
         "Cieľom bolo rozoznávať neznáme pomenované entity.",
     )
+
+def test_negative_answer_accepts_semantic_no_answer_wording() -> None:
+    question = {
+        "category": "unanswerable_missing_fact",
+        "expected_answer_contains": [],
+        "expected_source_urls": [
+            "https://example.test/student"
+        ],
+        "should_answer": False,
+    }
+
+    answer = (
+        "V dostupných dokumentoch ZP Wiki sa nepodarilo nájsť "
+        "informáciu o konkrétnom modeli GPU."
+    )
+
+    result = evaluate_answer(
+        question,
+        answer,
+        tool_called=True,
+    )
+
+    assert result["returned_no_answer"] is True
+    assert result["should_answer_ok"] is True
+    assert result["strict_pass"] is True
+
+
+def test_negative_answer_does_not_require_source_url() -> None:
+    question = {
+        "category": "unanswerable_missing_fact",
+        "expected_answer_contains": [],
+        "expected_source_urls": [
+            "https://example.test/student"
+        ],
+        "should_answer": False,
+    }
+
+    answer = (
+        "V dostupných dokumentoch ZP Wiki sa túto "
+        "informáciu nepodarilo spoľahlivo nájsť."
+    )
+
+    result = evaluate_answer(
+        question,
+        answer,
+        tool_called=True,
+    )
+
+    assert result["source_required_count"] == 0
+    assert result["source_url_score"] == 1.0
+    assert result["strict_pass"] is True
+
+def test_spacy_goal_matches_expanded_slovak_wording() -> None:
+    assert expected_phrase_matches(
+        "Zlepšiť presnosť modelu Spacy pre slovenčinu",
+        (
+            "Cieľom bolo zlepšiť presnosť modelu Spacy "
+            "pre spracovanie slovenčine."
+        ),
+    )
+
+
+def test_small_context_matches_short_context() -> None:
+    assert expected_phrase_matches(
+        "príliš maly kontext",
+        "Modely mali problém s príliš krátkou dĺžkou kontextu.",
+    )
+
+
+def test_slovak_hate_speech_matches_slovak_translation() -> None:
+    assert expected_phrase_matches(
+        "Slovak HATE speech",
+        "Úloha bola slovenská detekcia nenávistných prejavov.",
+    )
+
+
+def test_slovak_question_answering_matches_translation() -> None:
+    assert expected_phrase_matches(
+        "Slovak question answering",
+        "Išlo o slovenské vyhľadávanie odpovedí na otázky.",
+    )
+
+
+def test_named_matches_menovane_variant() -> None:
+    assert expected_phrase_matches(
+        "recognize unknown named entities",
+        "Cieľom bolo rozpoznať neznáme menované entity.",
+    )
+
+
+def test_multilingual_matches_many_languages_wording() -> None:
+    assert expected_phrase_matches(
+        "multilinguálna extrakcia trojíc z medicínskych dát",
+        (
+            "Projekt riešil extrakciu trojíc z lekárskych dát "
+            "v mnohých jazykoch."
+        ),
+    )
+
+
+def test_medical_data_matches_medical_content() -> None:
+    assert expected_phrase_matches(
+        "Triplet extraction from medical data",
+        (
+            "Témou bolo multilingválne extrahovanie trojíc "
+            "z lekárskeho obsahu."
+        ),
+    )