Simplified the json schema passed to the LLM, and implemented more forgiving error handling for LLM proposed corrections

tests/test_corrections.py

@@ -20,69 +20,95 @@ def test_apply_corrections_uses_threshold_and_sorts_chronologically():
     transcript = _transcript()
     corrections = [
         CorrectionCandidate(
-            segment_index=0,
-            speaker="Eric",
-            start=10.0,
-            end=11.0,
+            segment_id=0,
             original_text="I ask Chontia.",
             corrected_text="I ask Chauntea.",
             confidence=0.8,
         )
     ]
 
-    revised = apply_corrections(transcript, corrections, confidence_threshold=0.8)
+    result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
-    assert [segment.speaker for segment in revised] == ["Mike", "Eric"]
-    assert revised[1].text == "I ask Chauntea."
+    assert [segment.speaker for segment in result.transcript] == ["Mike", "Eric"]
+    assert result.transcript[1].text == "I ask Chauntea."
+    assert result.skipped == []
 
 
 def test_apply_corrections_ignores_below_threshold():
     transcript = _transcript()
     corrections = [
         CorrectionCandidate(
-            segment_index=0,
-            speaker="Eric",
-            start=10.0,
-            end=11.0,
+            segment_id=0,
             original_text="I ask Chontia.",
             corrected_text="I ask Chauntea.",
             confidence=0.79,
         )
     ]
 
-    revised = apply_corrections(transcript, corrections, confidence_threshold=0.8)
+    result = apply_corrections(transcript, corrections, confidence_threshold=0.8)
 
-    assert revised[1].text == "I ask Chontia."
+    assert result.transcript[1].text == "I ask Chontia."
+    assert result.skipped == []
 
 
-def test_apply_corrections_rejects_duplicate_targets():
+def test_apply_corrections_skips_duplicate_targets():
     transcript = _transcript()
-    correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=10.0,
-        end=11.0,
+    first = CorrectionCandidate(
+        segment_id=0,
         original_text="I ask Chontia.",
         corrected_text="I ask Chauntea.",
         confidence=0.8,
     )
+    second = CorrectionCandidate(
+        segment_id=0,
+        original_text="I ask Chontia.",
+        corrected_text="I ask Something Else.",
+        confidence=0.9,
+    )
 
-    with pytest.raises(AuditaValidationError):
-        apply_corrections(transcript, [correction, correction], confidence_threshold=0.8)
+    result = apply_corrections(transcript, [first, second], confidence_threshold=0.8)
+
+    assert result.transcript[1].text == "I ask Chauntea."
+    assert len(result.skipped) == 1
+    assert result.skipped[0].segment_id == 0
+    assert "duplicate" in result.skipped[0].reason
 
 
-def test_apply_corrections_rejects_mismatched_original_text():
+def test_apply_corrections_skips_mismatched_original_text():
     transcript = _transcript()
     correction = CorrectionCandidate(
-        segment_index=0,
-        speaker="Eric",
-        start=10.0,
-        end=11.0,
+        segment_id=0,
         original_text="Different text.",
         corrected_text="I ask Chauntea.",
         confidence=0.8,
     )
 
-    with pytest.raises(AuditaValidationError):
-        apply_corrections(transcript, [correction], confidence_threshold=0.8)
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
 
+    assert result.transcript[1].text == "I ask Chontia."
+    assert len(result.skipped) == 1
+    assert result.skipped[0].segment_id == 0
+    assert result.skipped[0].actual_text == "I ask Chontia."
+    assert "original_text" in result.skipped[0].reason
+
+
+def test_apply_corrections_skips_missing_segment_id():
+    transcript = _transcript()
+    correction = CorrectionCandidate(
+        segment_id=99,
+        original_text="Missing.",
+        corrected_text="Still missing.",
+        confidence=0.8,
+    )
+
+    result = apply_corrections(transcript, [correction], confidence_threshold=0.8)
+
+    assert [segment.text for segment in result.transcript] == ["Then Lyra.", "I ask Chontia."]
+    assert len(result.skipped) == 1
+    assert result.skipped[0].segment_id == 99
+    assert "does not exist" in result.skipped[0].reason
+
+
+def test_apply_corrections_rejects_invalid_threshold():
+    with pytest.raises(AuditaValidationError):
+        apply_corrections(_transcript(), [], confidence_threshold=1.1)