Reader

We are happy to announce the 1.0.0 Milestone 2 release of Spring AI.

This release brings significant enhancements and new features across various areas.

Key Focus: Observability

A major emphasis for this release has been on observability functionality, crucial for monitoring, debugging, and optimizing AI applications. Comprehensive observability features have been introduced for:

ChatClient (including Advisors)
Chat Models (OpenAI, Ollama, Mistral, Anthropic)
Embedding Models
Image Generation Models
Vector Stores

Note: Support for additional ChatModel implementation will be added in future releases.

Here are some samples showcasing this functionality to get started

Observability Demo from Christian Tzolov
Observability Demos from Thomas Vitale

Grazie mille (a thousand thanks ;) to Thomas Vitale for his significant efforts in contributing to the Observability feature!

Here is a screenshot showing the tracing of the following code

var response = chatClient.prompt()
	.user("How does Carina work?")
	.advisors(new QuestionAnswerAdvisor(vectorStore, SearchRequest.defaults()))
	.advisors(new PromptChatMemoryAdvisor(chatMemory))
	.call()
	.chatResponse();

Grafana view of Spring AI Trace

In the trace view, you can see the nesting of the ChatClient call which first passes through two advisors, the QuestionAnswerAdvisor that does an embedding request and a query to the PGVector database. Then it passes through the PromptChatMemoryAdvisor, which takes very little time, and then finally to the calling of the OpenAI gpt-4o model.

You can find all the details what metrics are available in the Observability Documentation

Core improvements

New: MarkdownDocumentReader for ETL pipelines
New: Cassandra-backed chat memory
New: Typesense testcontainer integration
ChatClient enhancements
- Function registration with explicit input type
- Dynamic filter expressions for QuestionAnswerAdvisor
- Customizable logger advisor
Improved API design of ResponseMetadata and the Message object hierarchy
Embedding format changed from List to float[]
Added JSON array handling to JsonReader
Options improvements
- ChatOptions: Enhanced metadata configuration for defined models
- EmbeddingOptions: Added model and dimensions options
- ImageOptions: Added style option
ChatOptions added to PromptTemplate create methods

ChatModel Enhancements

General improvements have been made across the chat model infrastructure classes. Key feature enhancements for specific models include:

Anthropic: Added Function calling support
Azure Open AI
- Function calling integration
- Updated to Azure OpenAI Client library 1.0 beta-10
- New: Transcription support
- Added image handling for byte arrays
- Added Multimodality support
- Added OpenAI API key and MS-Entra authentication options
Amazon Bedrock: Updated bedrockruntime version to 2.26.7
HuggingFace: Added Boot starter
Minimax
- Improved function calling
- Added Web search tool support
- Introduced sensitive information masking option
Mistral
- Added function calling support
- Documentation for using Mistral AI via Spring AI's OpenAI client
New: Moonshot chat model with function calling
Ollama
- Function calling support
- Batch embedding request functionality
OpenAI
- New: Moderation API support
- Updates for partially compliant OpenAI model providers
- Usage statistics for streaming chat responses
- Groq model provider integration with documentation
- Configurable URL paths for Chat and Embedding models
- Documentation for using Mistral AI and NVIDIA LLM API via Spring AI's OpenAI client
- Parallel function calling support
- Organization ID and Project ID authentication options
New: QianFan model for chat, embeddings, and image generation.
ONNX - Improved transformer embedding model
Google Vertex AI
- Add Embedding model support
- Updated Google Vertex SDK to 26.41.0
- Resolved parallel function calling
- Added system message support
- Added ResponeMimeType in chat options
- Added Google search retrieval tool support
ZhiPu: Added function calling support

Thanks for Geng Rong for the Moonshot, QianFan, and ZhiPu model support.

Vector Store improvements

Breaking Change: Default value for vector store configuration property initialize-schema changed to false.
Introduced BatchingStrategy for efficient document embedding operations
- Added TokenCountBatchingStrategy implementation
Cassandra: Added CassandraChatMemory
Chroma:
- Upgraded from use of RestTemplate to RestClient
- Integrated ChromaDB's CHROMA_SERVER_AUTHN_CREDENTIALS environment variable for Docker Compose and Testcontainers
Gemfire: Enhanced functionality
Milvus: Removed flush operations to prevent excessive segment creation
MongoDB: Improved efficiency using org.bson.Documents
Neo4j - Updated to latest Cypher-DSL
New: Opensearch VectorStore implementation Thanks Jemin Huh!
New: Oracle 23ai VectorStore implementation Thanks Loïc Lefèvre!
PGVector
- Introduced customizable schema, table, and index names
- Fixed IN/NOT IN filters for JSON queries
Pinecone
- Added configurable content and distance metadata fields
- Added AOT hints
Redis: Enabled use of standard RedisAutoConfiguration
New: Typesense support

General improvements

Many refactoring, bug fixing, documentation enhancements across the board by a wide range of contributors. If we haven’t gotten to your PR yet, we will, please be patient. Thanks to