Improve Ollama speaker detection with multi-sample strategy

- Sample up to 4 chunks of 5000 characters from different text positions
- Distribute samples evenly across beginning, middle, and end of text
- Deduplicate speakers across all samples for comprehensive detection
- Log individual sample results for debugging
- Handle short texts efficiently (single sample)
- Updated tests to verify multi-sample behavior

This ensures speakers introduced anywhere in long texts are detected,
not just those appearing in the first 2000 characters.

backend/test_ollama_service.py

@@ -24,6 +24,44 @@ async def test_ollama_service_custom_base_url():
     assert service.base_url == custom_url
 
 
+@pytest.mark.asyncio
+async def test_sample_text_short():
+    """Test sampling short text returns full text."""
+    service = OllamaService()
+    text = "Short text"
+    samples = service._sample_text(text, chunk_size=5000, num_samples=4)
+    
+    assert len(samples) == 1
+    assert samples[0] == text
+
+
+@pytest.mark.asyncio
+async def test_sample_text_long():
+    """Test sampling long text returns multiple samples."""
+    service = OllamaService()
+    # Create text with position markers so we can verify different samples
+    text = "".join([f"Position{i:05d} " for i in range(2500)])  # ~25k characters
+    samples = service._sample_text(text, chunk_size=5000, num_samples=4)
+    
+    assert len(samples) == 4
+    assert all(len(s) <= 5000 for s in samples)
+    # Samples should be from different parts - check they contain different position markers
+    assert "Position00000" in samples[0]  # First sample should have early positions
+    assert "Position00000" not in samples[-1]  # Last sample should have later positions
+
+
+@pytest.mark.asyncio
+async def test_sample_text_medium():
+    """Test sampling medium text returns appropriate number of samples."""
+    service = OllamaService()
+    text = "B" * 8000  # 8k characters
+    samples = service._sample_text(text, chunk_size=5000, num_samples=4)
+    
+    # Should get 2 samples (8000 / 5000 = 1.6, so 2 samples)
+    assert len(samples) >= 1
+    assert all(len(s) <= 5000 for s in samples)
+
+
 @pytest.mark.asyncio
 async def test_build_prompt():
     """Test prompt builder creates correct format."""
@@ -33,7 +71,7 @@ async def test_build_prompt():
     
     assert "speakers" in prompt
     assert "JSON" in prompt
-    assert text in prompt or text[:2000] in prompt
+    assert text in prompt
 
 
 @pytest.mark.asyncio
@@ -95,7 +133,7 @@ async def test_parse_response_invalid_json():
 
 @pytest.mark.asyncio
 async def test_detect_speakers_success():
-    """Test successful speaker detection."""
+    """Test successful speaker detection with short text."""
     service = OllamaService()
     text = "Alice: Hello!\nBob: Hi there!"
     
@@ -109,7 +147,32 @@ async def test_detect_speakers_success():
         mock_client.return_value.__aenter__.return_value.post = AsyncMock(return_value=mock_response)
         
         speakers = await service.detect_speakers(text)
-        assert speakers == ["Alice", "Bob"]
+        assert set(speakers) == {"Alice", "Bob"}
+
+
+@pytest.mark.asyncio
+async def test_detect_speakers_multiple_samples():
+    """Test speaker detection with long text requiring multiple samples."""
+    service = OllamaService()
+    text = "A" * 25000  # Long text requiring multiple samples
+    
+    # Mock responses for different samples
+    responses = [
+        MagicMock(json=lambda: {"response": '{"speakers": ["Alice", "Bob"]}'}),
+        MagicMock(json=lambda: {"response": '{"speakers": ["Charlie", "Bob"]}'}),
+        MagicMock(json=lambda: {"response": '{"speakers": ["Alice", "David"]}'}),
+        MagicMock(json=lambda: {"response": '{"speakers": ["Eve"]}'}),
+    ]
+    
+    for r in responses:
+        r.raise_for_status = MagicMock()
+    
+    with patch('httpx.AsyncClient') as mock_client:
+        mock_client.return_value.__aenter__.return_value.post = AsyncMock(side_effect=responses)
+        
+        speakers = await service.detect_speakers(text)
+        # Should deduplicate across all samples
+        assert set(speakers) == {"Alice", "Bob", "Charlie", "David", "Eve"}
 
 
 @pytest.mark.asyncio