…4132)

This commit addresses issues with OpenAI streaming responses where
finish_reason is an empty string ("") instead of null, which caused
Jackson deserialization failures and inconsistent metadata formatting.

Changes:
- Configure Jackson coercion in ModelOptionsUtils to handle empty
  string to enum conversion (empty string -> null)
- Fix getFinishReasonJson() method in OpenAiChatModel to return
  uppercase enum names for backward compatibility
- Fix null pointer exceptions in streaming processing by ensuring
  getFinishReasonJson() returns empty string instead of null
- Maintain finish_reason metadata as uppercase enum names
  ("STOP", "LENGTH") to preserve backward compatibility
- Add comprehensive test suite covering all edge cases
- Update streaming tests to expect correct uppercase finish reason values

Key fixes:
1. Empty string finish_reason now deserializes correctly without errors
2. Finish reason metadata maintains backward compatibility (uppercase enum names)
3. No more NullPointerException in streaming Map.of() calls
4. Proper handling of null, empty string, and valid finish_reason values

Jackson coercion configuration in ModelOptionsUtils now converts empty
strings to null for ALL enum deserialization across Spring AI. Analysis
shows this change is safe and beneficial for other providers:
- Anthropic, Ollama use String fields (unaffected by enum coercion)
- ZhiPuAI, MiniMax, MistralAI, DeepSeek use enum fields but have no
  documented empty string issues
- Coercion improves robustness by gracefully handling malformed responses
- No breaking changes expected for existing provider implementations

Testing:
- Added OpenAiStreamingFinishReasonTests with 6 comprehensive test cases
- Fixed existing metadata tests to maintain backward compatibility
- All tests pass, covering JSON deserialization and streaming scenarios
- Validates both error handling and correct behavior
- Added tests for changes to ModelOptionsUtils using enum coersion

This ensures robust handling of OpenAI's empty string finish_reason edge
case while maintaining backward compatibility and improving overall
system resilience for malformed API responses.

Fixes #1358

Signed-off-by: Mark Pollack <[email protected]>
(cherry picked from commit 653918d)

…pring-projects#4132)

This commit addresses issues with OpenAI streaming responses where
finish_reason is an empty string ("") instead of null, which caused
Jackson deserialization failures and inconsistent metadata formatting.

Changes:
- Configure Jackson coercion in ModelOptionsUtils to handle empty
  string to enum conversion (empty string -> null)
- Fix getFinishReasonJson() method in OpenAiChatModel to return
  uppercase enum names for backward compatibility
- Fix null pointer exceptions in streaming processing by ensuring
  getFinishReasonJson() returns empty string instead of null
- Maintain finish_reason metadata as uppercase enum names
  ("STOP", "LENGTH") to preserve backward compatibility
- Add comprehensive test suite covering all edge cases
- Update streaming tests to expect correct uppercase finish reason values

Key fixes:
1. Empty string finish_reason now deserializes correctly without errors
2. Finish reason metadata maintains backward compatibility (uppercase enum names)
3. No more NullPointerException in streaming Map.of() calls
4. Proper handling of null, empty string, and valid finish_reason values

Jackson coercion configuration in ModelOptionsUtils now converts empty
strings to null for ALL enum deserialization across Spring AI. Analysis
shows this change is safe and beneficial for other providers:
- Anthropic, Ollama use String fields (unaffected by enum coercion)
- ZhiPuAI, MiniMax, MistralAI, DeepSeek use enum fields but have no
  documented empty string issues
- Coercion improves robustness by gracefully handling malformed responses
- No breaking changes expected for existing provider implementations

Testing:
- Added OpenAiStreamingFinishReasonTests with 6 comprehensive test cases
- Fixed existing metadata tests to maintain backward compatibility
- All tests pass, covering JSON deserialization and streaming scenarios
- Validates both error handling and correct behavior
- Added tests for changes to ModelOptionsUtils using enum coersion

This ensures robust handling of OpenAI's empty string finish_reason edge
case while maintaining backward compatibility and improving overall
system resilience for malformed API responses.

Auto-cherry-pick to 1.0.x
Fixes spring-projects#1358

Signed-off-by: Mark Pollack <[email protected]>

…pring-projects#4132)

This commit addresses issues with OpenAI streaming responses where
finish_reason is an empty string ("") instead of null, which caused
Jackson deserialization failures and inconsistent metadata formatting.

Changes:
- Configure Jackson coercion in ModelOptionsUtils to handle empty
  string to enum conversion (empty string -> null)
- Fix getFinishReasonJson() method in OpenAiChatModel to return
  uppercase enum names for backward compatibility
- Fix null pointer exceptions in streaming processing by ensuring
  getFinishReasonJson() returns empty string instead of null
- Maintain finish_reason metadata as uppercase enum names
  ("STOP", "LENGTH") to preserve backward compatibility
- Add comprehensive test suite covering all edge cases
- Update streaming tests to expect correct uppercase finish reason values

Key fixes:
1. Empty string finish_reason now deserializes correctly without errors
2. Finish reason metadata maintains backward compatibility (uppercase enum names)
3. No more NullPointerException in streaming Map.of() calls
4. Proper handling of null, empty string, and valid finish_reason values

Jackson coercion configuration in ModelOptionsUtils now converts empty
strings to null for ALL enum deserialization across Spring AI. Analysis
shows this change is safe and beneficial for other providers:
- Anthropic, Ollama use String fields (unaffected by enum coercion)
- ZhiPuAI, MiniMax, MistralAI, DeepSeek use enum fields but have no
  documented empty string issues
- Coercion improves robustness by gracefully handling malformed responses
- No breaking changes expected for existing provider implementations

Testing:
- Added OpenAiStreamingFinishReasonTests with 6 comprehensive test cases
- Fixed existing metadata tests to maintain backward compatibility
- All tests pass, covering JSON deserialization and streaming scenarios
- Validates both error handling and correct behavior
- Added tests for changes to ModelOptionsUtils using enum coersion

This ensures robust handling of OpenAI's empty string finish_reason edge
case while maintaining backward compatibility and improving overall
system resilience for malformed API responses.

Auto-cherry-pick to 1.0.x
Fixes spring-projects#1358

Signed-off-by: Mark Pollack <[email protected]>

…pring-projects#4132)

This commit addresses issues with OpenAI streaming responses where
finish_reason is an empty string ("") instead of null, which caused
Jackson deserialization failures and inconsistent metadata formatting.

Changes:
- Configure Jackson coercion in ModelOptionsUtils to handle empty
  string to enum conversion (empty string -> null)
- Fix getFinishReasonJson() method in OpenAiChatModel to return
  uppercase enum names for backward compatibility
- Fix null pointer exceptions in streaming processing by ensuring
  getFinishReasonJson() returns empty string instead of null
- Maintain finish_reason metadata as uppercase enum names
  ("STOP", "LENGTH") to preserve backward compatibility
- Add comprehensive test suite covering all edge cases
- Update streaming tests to expect correct uppercase finish reason values

Key fixes:
1. Empty string finish_reason now deserializes correctly without errors
2. Finish reason metadata maintains backward compatibility (uppercase enum names)
3. No more NullPointerException in streaming Map.of() calls
4. Proper handling of null, empty string, and valid finish_reason values

Jackson coercion configuration in ModelOptionsUtils now converts empty
strings to null for ALL enum deserialization across Spring AI. Analysis
shows this change is safe and beneficial for other providers:
- Anthropic, Ollama use String fields (unaffected by enum coercion)
- ZhiPuAI, MiniMax, MistralAI, DeepSeek use enum fields but have no
  documented empty string issues
- Coercion improves robustness by gracefully handling malformed responses
- No breaking changes expected for existing provider implementations

Testing:
- Added OpenAiStreamingFinishReasonTests with 6 comprehensive test cases
- Fixed existing metadata tests to maintain backward compatibility
- All tests pass, covering JSON deserialization and streaming scenarios
- Validates both error handling and correct behavior
- Added tests for changes to ModelOptionsUtils using enum coersion

This ensures robust handling of OpenAI's empty string finish_reason edge
case while maintaining backward compatibility and improving overall
system resilience for malformed API responses.

Auto-cherry-pick to 1.0.x
Fixes spring-projects#1358

Signed-off-by: Mark Pollack <[email protected]>
Signed-off-by: 家娃 <[email protected]>