- Add Pixtral models (PIXTRAL and PIXTRAL_LARGE) to MistralAiApi.ChatModel
- Update MistralAiChatClientIT to use Pixtral model for testing
- Add new Ollama models:
* QWEN_2_5_7B
* LLAMA3_2_VISION_11b
* LLAMA3_2_VISION_90b
Fixes: #1753https://github.com/spring-projects/spring-ai/issues/1753
- Add character-based truncation (max 2048 chars) for Cohere embedding requests
- Support both START and END truncation strategies
- Add unit tests verifying truncation behavior for both strategies
Truncation is applied before sending requests to Bedrock API to avoid
ValidationException when text exceeds maximum length. The END strategy
(default) keeps the first 2048 characters while START keeps the last
2048 characters.
- Replace hardcoded "tool_use" with StopReason enum value
- Add tests for token usage aggregation with tool calls
- Add handling for null response metadata
- Modify stream method to support recursive tool call handling
- Update token tracking and metadata merging for streamed responses
- Improve token usage calculation for tool use events
- Update test cases to handle new response processing
- Modify call method to support recursive tool call handling
- Add support for cumulative token tracking across tool call iterations
- Introduce internal call method to track and aggregate token usage
- Merge previous chat response tokens with current response tokens
Resolves#1743
- Add TOOL_CALL_HISTORY constant to store tool call history
- Extend ToolContext with getToolConversationHistory method
- Include tool call history in tool context during function execution
- Add test coverage for tool call history verification
Resolves#1202
The commit restructures OpenAI token usage tracking by:
- Adding audio_tokens support in PromptTokensDetails
- Deprecating individual token getter methods in favor of consolidated records
- Introducing new PromptTokensDetails and CompletionTokenDetails records
- Updating tests to reflect the new structure
Resolves#1369 , #1720
Add support for no-argument Supplier and single-argument Consumer function
callbacks in the Spring AI core module. This enhancement allows:
- Registration of Supplier<O> callbacks with no input (Void) type
- Registration of Consumer<I> callbacks with no output (Void) type
- Support for Kotlin Function0 (equivalent to Java Supplier)
- Handle empty properties for Void input types in schema generation
- Enhance FunctionCallback builder to support Supplier/Consumer patterns
Additional changes:
- Add test coverage for both Supplier and Consumer callbacks in various scenarios
- Enhance TypeResolverHelper to support Consumer input type resolution
- Support lambda-style function declarations for improved ergonomics
- Add test cases for void input/output handling in OpenAI chat model
- Include examples of function calls without return values
- Add support for parameterless functions through Supplier interface
Add comprehensive documentation for the FunctionCallback API:
- Overview of the interface and its key methods
- Builder pattern usage with function and method invocation approaches
- Examples for different function types (Function, BiFunction, Supplier, Consumer)
- Best practices and common pitfalls
- Schema generation and customization options
Resolves#1718 , #1277 , #1118, #860
Introduces a simplified, type-safe builder pattern for function callbacks to
improve developer experience and code reliability. The new hierarchical API
separates concerns between direct function invocation and method reflection,
while providing better compile-time safety.
This change deprecates the older FunctionCallbackWrapper in favor of a more
intuitive FunctionCallback.Builder that better handles generic types via
ParameterizedTypeReference. It also adds automatic function description
generation as a fallback when none is provided, though explicit descriptions
are still recommended.
The update standardizes function callback handling across all AI model
implementations (OpenAI, Ollama, Minimax, etc.) and improves response
handling with configurable converters.
Core API Enhancements:
- New Builder Interface: Replaced FunctionCallbackWrapper.builder() with
FunctionCallback.builder(), introducing a hierarchical approach that improves
customization and type safety.
- Specialized Builders: Introduced FunctionInvokerBuilder for direct Function/BiFunction
implementations and MethodInvokerBuilder for reflection-based invocations.
- Generic Type Support: Added ParameterizedTypeReference for better handling of generic parameters.
- Unified Method Definition: Merged method() and argumentTypes() into a single method() call
for simplicity and type safety.
- Automatic Descriptions: Implemented auto-generation of function descriptions, with warnings
to encourage explicit descriptions.
- Configurable Response Converters: Enhanced response handling with support for custom
converters, reducing unnecessary JSON conversions.
Architecture Improvements:
- Established common Builder interface for shared properties
- Separated function object handling from constructor
- Added method-specific configuration (name, arg types, target)
- Added JSON schema generation support for ResolvableType
- Moved to standardized schema types across AI providers
- Set OPEN_API_SCHEMA as default for Vertex AI Gemini
Builder Pattern Standardization:
- Standardized builder method ordering across implementations
- Moved function() call after description() for consistency
- Improved function callback configuration with unified patterns
- Enhanced error handling and validation in DefaultFunctionCallbackBuilder
Deprecations:
- FunctionCallbackWrapper.Builder replaced by DefaultFunctionCallbackBuilder
- Removed CustomizedTypeReference in favor of ParameterizedTypeReference
- Deprecated older ChatClient API methods for function handling
Testing & Documentation:
- Updated all AI model implementations (OpenAI, Ollama, Minimax, Moonshot, ZhiPuAI)
- Added comprehensive integration tests for static/instance methods
- Added integration tests for auto-generated descriptions
- Updated documentation to reflect new builder pattern usage
- Added Kotlin extension for inputType() support
Co-authored-by: Sébastien Deleuze <sebastien.deleuze@broadcom.com>
- Remove deprecated types,methods and references
- Remove usage of "Generation(String text)" and replace with
"Generation(AssistantMessage)"
- Remove MiniMaxApi,MootShotApi,OpenAiApi,ZhiPuAiApi
ChatCompletionFinishReason's FUNCTION_CALL
- Remove OllamaApi's deprecated types
- Remove deprecated constructors from PostgresMlEmbeddingModel
- Remove deprecated constructor from Media
- Remove tokenNames usage from antlr4 FiltersLexer and FiltersParser
- Remove deprecated methods from CassandraVectorStoreProperties
- Remove deprecated constructor and static config class from QdrantVectorStore
- Minor cleanups on removing deprecated models and versions
- Remove old name for mixtral models
Resolves#1599
- Update GenerateResponse content schema type to array at openapi.json
- Use CompatGenerateRequest instead of GenerateRequest for the TextGenerationInference API Request
Signed-off-by: jitokim <pigberger70@gmail.com>
Due to changes in Chroma v0.5.13 (chroma-core/chroma#2880), the delete
operation no longer returns values. This impacts our ChromaVectorStore's delete
functionality.
Now we need to check the HTTP status code instead to properly verify if the
delete operation succeeded.
Test suite has been updated.
Resolves#1529
Query Analysis
* Introduce Query Analysis Module
* Define QueryTransformer API and TranslationQueryTransformer implementation
* Define QueryExpander API and MultiQueryExpander implementation
* Support QueryTransformer in RetrievalAugmentationAdvisor (support for QueryExpander will be in the next PR together with the needed DocumentFuser API).
Improvements
* Refine Retrieval and Augmentation Modules for increased robustness
* Expand test coverage for both modules
* Define clone() method for ChatClient.Builder
Tests
* Introduce “spring-ai-integration-tests” for full-fledged integration tests
* Add integration tests for RAG modules
* Add integration tests for RAG advisor
Query Analysis
* Introduce Query Analysis Module
* Define QueryTransformer API and TranslationQueryTransformer implementation
* Define QueryExpander API and MultiQueryExpander implementation
* Support QueryTransformer in RetrievalAugmentationAdvisor (support for QueryExpander will be in the next PR together with the needed DocumentFuser API).
Improvements
* Refine Retrieval and Augmentation Modules for increased robustness
* Expand test coverage for both modules
* Define clone() method for ChatClient.Builder
Tests
* Introduce “spring-ai-integration-tests” for full-fledged integration tests
* Add integration tests for RAG modules
* Add integration tests for RAG advisor
Relates to #gh-1603
Signed-off-by: Thomas Vitale <ThomasVitale@users.noreply.github.com>
This change enables more flexible integration between Spring AI and LLM function
calling capabilities while maintaining type safety and ease of use.
- Add new MethodFunctionCallback class to support method invocation via reflection
- Supports both static and non-static method calls
- Handles multiple parameter types including primitives, objects, collections
- Supports empty parameters and empty response
- Auto-generates JSON schema from method parameters
- Special handling for ToolContext parameters
- Builder pattern for easy configuration
- Add comprehensive unit tests for MethodFunctionCallback
- Add integration tests for MethodFunctionCallback with both Anthropic and OpenAI clients
- Add jackson-module-jsonSchema dependency
- Modify FunctionCallback to check for empty tool context
Testing coverage includes:
- Static method invocation scenarios
- Non-static method calls with various parameter types
- Void return type methods
- Complex parameter types (enums, records, lists)
- Tool context handling
- Error cases and validation
Add MethodFunctionCallback reference docs
This change simplifies how we manage Ollama containers in tests by moving
from manual toggles to environment variables for better control. Instead of
scattered container configuration, we now have:
- OLLAMA_WITH_REUSE: Toggle reuse of existing containers between tests
- OLLAMA_TESTS_ENABLED: Control test execution globally
The motivation is to make tests more reliable and easier to maintain.
Previously, developers had to modify code to run tests locally vs CI. Now
they can control this via environment variables.
We also introduce thread-safe API access and consistent default settings
across all test classes, removing duplicated configuration and potential
resource leaks.
This makes the test infrastructure more maintainable and provides clearer
separation between local development and CI environments.
Checkstyle fixes.
Make buildOllamaApiWithModel in BaseOllamaIT public and the related test changes.
- Verify that the BeanOutputConverter converts with the right order as specified in the JSON schema
Update the BeanOutputConverterTest's test case to tweak the order for validation
The changes include:
- Refactor the emitting of next, error, and complete events in the Bedrock stream handling to use a default EmitFailureHandler that retries for 10 seconds before failing.
This helps improve the error resilience of the stream processing.
- Disable a couple of integration tests related to the COHERE_COMMAND_V14 model, as that model version is no longer supported.
- Adjust some configuration options in the Jurassic2 chat model integration test.
Also resolves#1679
XFY Spark and possibly other vendors implement the OpenAI interface but return
tool_calls as a single object rather than an array. Add @JsonFormat annotation
to handle both array and single object responses in ChatCompletionMessage,
making Spring AI more broadly compatible with OpenAI API implementations.
Only ChatCompletionMessage.java required modification to support this alternate
format while maintaining backward compatibility with the standard OpenAI API
response structure.
Co-authored-by: jito <pigberger70@gmail.com>
- Extracts ResponseFormat from being a nested record in OpenAiApi to
a dedicated class with builder pattern support.
- Resolve the issue with constructor bindings for the Boog property
binding.
- Re-enables previously disabled response format integration tests.
- Add checkstyle changes
- Add schema field in ResponseFormat and set jsonSchema via the setter for schema,
this way schema set via a Boot property also sets the correct JsonSchema
- Add default constructors in ResponseFormat and JsonSchema
Resolves#1681
Adds Oracle Cloud Infrastructure (OCI) Generative AI's Cohere chat model support
to expand Spring AI's cloud provider capabilities. This allows developers to use
OCI's managed Cohere models through both dedicated and on-demand serving modes.
The integration provides auto-configuration for simple setup while allowing full
customization of model parameters through OCICohereChatOptions. Teams can now
use OCI's Cohere models alongside other providers in Spring AI applications.
This change complements the existing OCI embedding support, offering a complete
set of GenAI capabilities for Oracle Cloud users.
Signed-off-by: Anders Swanson <anders.swanson@oracle.com>
OpenAI's API returns additional token usage metrics that provide deeper
insight into API consumption. This adds support for:
- acceptedPredictionTokens: Tokens from accepted model predictions
- audioTokens: Tokens used for audio processing
- rejectedPredictionTokens: Tokens from rejected model predictions
These fields help track resource utilization and costs more accurately
by breaking down token usage by type. Added @JsonIgnoreProperties to
maintain compatibility with future OpenAI API additions.
Fixes warning logging in RetryUtils.SHORT_RETRY_TEMPLATE to reduce noise
in test output.
The DEFAULT_RETRY_TEMPLATE uses exponential backoff starting at 2s with max 3min
delay between retries, making tests run unnecessarily long. This change
introduces SHORT_RETRY_TEMPLATE with fixed 100ms backoff to speed up test
execution while preserving the same retry behavior and error handling.
- Add a new test case for testing function call with advisor in BedrockConverseChatClientIT
- Add a new test case for testing tool proxy function call in OpenAiChatClientProxyFunctionCallsIT
- Remove unused model property from BedrockConverseProxyChatProperties
- Disable project-wide Checkstyle checks to unblock development
- Add documentation for enabling Checkstyle locally
- Fix remaining checkstyle violations in current codebase
Fixes#1669
Add @JsonIgnore annotation to the jsonSchema field in Function classes across
multiple AI provider APIs (MiniMax, MistralAI, OpenAI, ZhiPuAI) to exclude
it from JSON serialization.
