⭐ New Features

• Expose the version-checked append on SessionService #7314
• Add server support for the MCP 2026-07-28 ("modern") protocol revision #7284
• Implement OllamaDecisionModel #7261
• Add DecisionToolIndex for tool search #7253
• Add decision-based RAG document post-processors #7252
• Add SystemOneDecisionModel #7249
• Stream tool call argument deltas #7233
• Introduce DecisionModel API #7205
• Upgrade Anthropic Java SDK to 3.4.0 #7306
• Add upsert and query-embedding search to Chroma #7230
• Add upsert and query-embedding search to MariaDB #7235
• Add upsert and query-embedding search to MongoDB Atlas #7272
• Add upsert and query-embedding search to Typesense #7303
• Add upsert and query-embedding search to Neo4j #7318
• Include cached tokens in Anthropic and Bedrock prompt tokens #7268

vector store

❤️ Contributors

Thank you to all the contributors who worked on this release:

@ThomasVitale, @dimitarproynov, @sobychacko, @iuliiasobolevska, @Ajax-1 , and @tzolov

For upgrading from 2.1.0-M1 to 2.1.0-RC1, please refer to the upgrade notes here.

⭐ New Features

• Deprecate ChatMemory in favor of session memory #7186
• Add upsert and query-embedding search to OpenSearch #7177
• Fail fast on unsupported upsert and vector search #7176
• Add session memory: an event log with turn-aware compaction #7161
• Add tool call confirmation to tool calling #7157
• Migrate DeepSeek chat model to message parts #7128
• Migrate OpenAI chat model to message parts #7107
• Fail clearly on out-of-range OpenAI Responses usage counts #7052
• Support for Text Chunking with Overlap in TokenTextSplitter #2123

🔨 Dependency Upgrades

• Upgrade to Kotlin 2.4.21 #7168
• Upgrade Tika to 4.0.0 in spring-ai-tika-document-reader #7068

❤️ Contributors

Thank you to all the contributors who worked on this release:

@56wj, @Ashfaqbs, @OreOreDa, @Rainmemery, @dimitarproynov, @iuliiasobolevska, @sedhossein, @sobychacko, @ss-bhatt, @stack3mpty, and @tzolov

For upgrading from 2.0.1 to 2.0.2, please refer to the upgrade notes here.

⭐ New Features

• Upgrade the Anthropic Java SDK to 2.72.0 #7164
• Add observability support for TextToSpeechModel, TranscriptionModel and StabilityAiImageModel #7141
• Configure Redis vector store metadata fields #6870
• Add customizers for stateless MCP servers #6539

🐞 Bug Fixes

• Fix numeric filters in Cassandra vector store #7117
• Fix StructuredOutputValidationAdvisor repeating tool call requests #7100
• Handle Ollama responses without a message #7066
• Return final answer first in AnthropicChatModel #7001
• Honor disabled reasoning in tool history #6984
• Include thinking text in Anthropic streaming chunks #6982
• Fix Neo4j chat memory index properties #6967
• Fix invalid JSON in Bedrock tool-call arguments #6963
• Backfill additionalProperties: false in OpenAI strict tool schemas #6865
• Restore honoring the configured OpenAI timeout #6837
• Fix JsoupDocumentReader dropping relative link URLs #6803
• Close the resource stream in TextReader #6797
• Fix tool call extraction in GoogleGenAiChatModel for mixed content responses #5468

📔 Documentation

• Document Anthropic server-side spans in traces #7165
• Correct stale API names in AI metadata docs #7032
• docs: document OpenAI reasoning effort property #6943
• Document MCP client scope and session boundaries #6820

🔨 Dependency Upgrades

• Upgrade OpenAI Java SDK to 4.80.0 #7170
• Upgrade AWS SDK and Bedrock runtime to 2.55.14 #7169

❤️ Contributors

Thank you to all the contributors who worked on this release:

@CodeTrainerMan, @CryoThrust, @JamesBLewis, @Lubaoshuai, @Rainmemery, @ashakirin, @chabinhwang, @chengwudi1, @chensishang, @dimitarproynov, @dlwldn30, @flex-ilsub, @herder, @ikaitist, @iuliiasobolevska, @jhpark1227, @jianran, @kezhenxu94, @lazer-dev, @martin-grofcik, @nicolaskrier, @pengjonas, @pengmoubuaixuexi, @rvullriede, @ryujungkyun, @seo92js, @sharanrenukeshbharmshetty, @sobychacko, @tzolov, @vinayreddykalluri, @wantaekchoi, @xujiantop-crypto, and @zxuhan

For upgrading from 2.0.1 to 2.1.0-M1, please refer to the upgrade notes here.

⭐ New Features

• Add OpenAI Responses API support #7006
• Add VectorStore.upsert for pre-computed embeddings #7002
• Introduce MessagePart in AbstractMessage #6997

🔨 Dependency Upgrades

• Upgrade Spring Boot to 4.2 #7003

❤️ Contributors

Thank you to all the contributors who worked on this release:

@dimitarproynov and @sobychacko

For upgrading from 2.0.0 to 2.0.1, please refer to the upgrade notes here.

For the detailed reference documentation on 2.0.1, please refer here

⭐ New Features

• Harden Ollama integration tests against flakiness #6776
• Retire deprecated Mistral AI chat models #6772
• Document additional 2.0.1 breaking changes #6771
• Support audio streaming in OpenAiAudioSpeechModel #6733
• Add configurable tool call limits #6726
• Enrich OpenAI transcription options and responses #6697
• Update Couchbase vector store to use Spring Boot-managed client #6583
• Remove explicit protobuf-java version properties #6552
• Add ToolChoice support for Google GenAI chat model #6531
• Remove fastjson2 dependency from spring-ai-azure-store #6508
• Refactor Media builder to use typed data overloads #6481
• Validate TokenTextSplitter builder arguments #6452
• Support page ranges in PagePdfDocumentReader #6445
• Revise DeepSeekApi #6428
• GraalVM for mcp annotations doesn't work in Spring AI 2.0.0 #6427
• Rename spring-ai-autoconfigure-model-chat-memory-redis #6416
• Restore PDF media mapping in OpenAiChatModel #6410
• Google Image generation support #6408
• Remove unused field maxResults #6344
• Confusing WARN log while AWS region resolving #5108
• SpringAI does not follow the AWS SDK's default region resolution rules #823

🐞 Bug Fixes

• Reserve Redis chat memory timestamps atomically #6784
• Fix RedisChatMemoryRepository clear() #6783
• Escape metadata values in Redis chat memory queries #6765
• Validate resolved resource path #6764
• Introduce maxSessions and sessionIdleTimeout #6760
• Fix OpenAiAudioTranscriptionModel leaking the stream on cancellation #6759
• Merge java-compile compilerArgs instead of overriding #6756
• Default OpenAI tool-calling strict mode to false #6755
• Use temporary directory for cache #6754
• Handle malformed pdf in PdfDocumentReader #6753
• Make tool resolution fallback configurable #6751
• Use full length hash #6746
• Correct ToolContext parameter detection #6744
• Return a single Generation on ToolCallLimitExceededException #6742
• Fix ToolCallingAdvisor streaming loop dropping doBeforeStream mutations #6737
• Add instructions option to OpenAiAudioSpeechModel #6731
• Apply MCP HTTP client request customizer beans #6716
• Close OpenAI stream response on cancellation #6656
• Validate text length to prevent StringIndexOutOfBoundsException for MarkdownCodeBlockCleaner #6645
• Fix AnthropicChatOptions.timeout(Duration) being ignored outside model construction #6605
• Close native resources in TransformersEmbeddingModel #6572
• Fix OllamaChatModel prompt options not inheriting configured model #6567
• Use Locale.ROOT for String uppercase/lowercase operations on wire/protocol values #6566
• Handle null fields in Mistral AI moderation Categories and CategoryScores #6565
• Fix Azure OpenAI authentication failure behind a TLS-inspecting proxy #6553
• Fix OpenSearchVectorStore document deserialization #6550
• Fix group precedence in MilvusFilterExpressionConverter #6543
• Fix strict mode payload placement in OpenAiChatModel #6540
• Align @McpTool exception handling with @Tool contract #6534
• Fix MCP stdio connections without args #6533
• Fix Milvus metadata integer values deserialized as Long instead of Double #6517
• Fix OpenAI stream handling for empty tool_calls #6504
• Fix shared mutable state in chat options builder clone() #6498
• Fix tool-search reference extraction for parallel calls #6497
• Fix incorrect assertions in CouchbaseSearchVectorStore.Builder #6490
• Fix shared mutable state in AnthropicChatOptions.Builder.clone() #6487
• Fix null parent for tool call observations in blocking mode #6475
• Fix Milvus filters with custom metadata field #6470
• Add tool-search advisor modules to BOM #6464
• Use model-independent assertions for tool-call token usage #6461
• Fix tool-search advisor auto-config ordering #6458
• There is a bug in the openai streaming mode #6441
• Validate PgVectorStore table schema after initialization #6438
• Accumulate tool calling advisor usage #6424
• Fix MarkdownCodeBlockCleaner discarding content of single-line fenced code blocks #6423
• Merge streaming tool call deltas by index in OpenAiChatModel #6381
• reasoningContent in metadata is always null when calling DeepSeek model via OpenAI API #6375
• Fix genai thinking level validation #6339
• Fix Microsoft Foundry URL path mode for proxy hosts #6180
• Perform Optional unwrapping in OpenAiChatModel#buildGeneration #6009
• Fix Streamable HTTP WebClient transport to accept SSE frames without event field #5783

📔 Documentation

• Fix inaccurate Content interface Javadoc #6704
• Document conversation ID scoping for chat memory #6651
• Fix OpenAI audio documentation title and broken link #6606
• Restructure structured-output documentation #6495
• Restore the CosmosDBChatMemoryRepository section #6491
• Fix typos in McpTool JavaDoc comments #6449
• Restructure tool calling reference docs #6432
• Remove ToolSearchToolCallingAdvisor for upgrade notes [#6421](https://g...

For upgrading from 1.1.x to 2.0.0, please refer to the upgrade notes here.

For the detailed reference documentation on 2.0.0, please refer here

⭐ New Features

• Update Google GenAI models #6406
• Use only Jackson 2 in OpenAiChatModel #6392
• Make org.springframework.ai.image.observation null-marked #6388
• Polish Prompt #6387
• Preserve OpenAI tool call additional properties #6365

🐞 Bug Fixes

• Filter unsupported tool messages in CassandraChatMemoryRepository #6400
• Filter unsupported tool messages in MongoChatMemoryRepository #6399
• Filter unsupported tool messages in JdbcChatMemoryRepository #6398
• Remove streamToolCallResponses from advisor builders #6391
• Add missing promptCacheKey parameter in OpenAiChatOptions #6380

📔 Documentation

• Update reference docs to reflect 2.0 API changes #6405
• Add more guides #6401
• Introduce CLAUDE.md and AGENTS.md #6376
• Consolidate upgrade notes for 2.0.0 #6333

🔨 Dependency Upgrades

• Upgrade to MCP SDK 2.0.0 #6385
• Upgrade to Spring Boot 4.1.0 #6329

❤️ Contributors

Thank you to all the contributors who worked on this release:

@ericbottard, @guanxuc, @ilayaperumalg, @sdeleuze, @sdudzin, @tzolov, and @ultramancode

🐞 Bug Fixes

• Add missing values to ZhiPuAiApi.ChatCompletionFinishReason #5532

📔 Documentation

• Add a rate-limit metadata section to anthropic-chat.adoc #6175

🔨 Dependency Upgrades

• Upgrade Spring Boot to 3.5.15 #6393
• Upgrade MCP SDK to 0.18.3 #6362

❤️ Contributors

Thank you to all the contributors who worked on this release:

@demo-zexuan and @sobychacko

⭐ New Features

• Deprecate Pixtral models #6148

🔨 Dependency Upgrades

• Upgrade Spring Boot to 3.5.15 #6394
• Upgrade MCP SDK to 0.18.3 #6361

⭐ New Features

• Make Anthropic HTTP client configurable #6354
• Restore compatibility with Spring Framework < 7.0.4 #6334
• Make OpenAI HTTP Client configurable #6294

🐞 Bug Fixes

• Fix BedrockProxyChatModel model option handling #6370
• Fix spring-ai-autoconfigure-model-bedrock-ai dependencies #6368
• 2.0.0-RC1 regression: additional dependency is needed for Bedrock SDK #6367
• Always auto-register ToolCallingAdvisor to support runtime-injected tools #6349
• Restore options replacing instead of merging in ChatModel #6336
• Replay reasoning_content in OpenAI assistant history #6296
• Fix OllamaChatModel dropping thinking field in multi-turn conversation history #6062
• Binding failure for spring.ai.ollama.chat.think #4853

📔 Documentation

• Document native structured output limitations for Ollama and OpenAI #6363
• Rewrite CONTRIBUTING.adoc and convert it to CONTRIBUTING.md #6352
• Fix tool callback example in user-controlled tool execution docs #6324

❤️ Contributors

Thank you to all the contributors who worked on this release:

@ThomasVitale, @ericbottard, @guanxuc, @ilayaperumalg, @sdeleuze, @sdudzin, @suryateja-g13, @tzolov, and @uc4w6c

⭐ New Features

• Add turn-boundary snapping to MessageWindowChatMemory eviction #6312
• Finish to remove default values from models #6311
• Move advisors to their proper modules #6309
• Remove SpringBeanToolCallbackResolver and toolNames() API #6301
• Streamline OpenAiChatModel implementations #6295
• Remove ToolSpec consumer API from ChatClient #6292
• Remove Minimax dedicated support in favor of Anthropic one #6290
• Remove internalToolExecutionEnabled #6289
• Rename N() to n() in options builders #6285
• Refine JdbcChatMemoryRepositoryDialect to use default method #6283
• Replace SLF4J by org.apache.commons.logging.LogFactory #6280
• Remove internal tool execution from MiniMaxChatModel #6279
• Remove internal tool execution from BedrockProxyChatModel #6272
• fix(core): allow tool callbacks and providers directly in ChatClient tools() #6263
• Remove internal tool execution from MistralAiChatModel #6260
• Remove internal tool execution from DeepSeekChatModel #6258
• Remove internal tool execution from OpenAiChatModel #6257
• Remove internal tool execution from OllamaChatModel #6256
• Remove duplicate builder overrides in OpenAiChatOptions #6255
• Refactor tool search by consolidating into a single module #6254
• Remove internal tool execution from AnthropicChatModel #6252
• Refactor toOptions method for vendor-specific autoconfiguration properties #6248
• Make equals parameter nullable #6239
• Remove unused GoogleGenAiChatOptions.TransportType #6238
• GH-6211: Avoid duplicate chat memory in tool prompts #6228
• GH-6206: Update tool callback guidance #6214
• Update Mistral AI chat models #6210
• Deprecate and rename ChatClient customizer #6204
• Refine Spring Boot dependencies #6200
• Simplify ToolCallAdvisor stream aggregation by removing publish() multicast #6195
• Add DeepSeek V4 chat model constants #6194
• Add EntityParamSpec to ChatClient entity() methods for per-call structured output configuration #6165
• Remove toolNames() API and SpringBeanToolCallbackResolver — use explicit ToolCallback beans instead #6154
• Retire Pixtral Large model #6149
• Refactor JSON utilities #6135
• Use .class reference for autoconf ordering #6119
• Add Tool Search Tool advisor for on-demand tool discovery #5909
• Allow accessing "timestamp" field from spring_ai_chat_memory table programatically when using JdbcChatMemoryRepository #5168

🐞 Bug Fixes

• fix: place memory advisors outside ToolCallingAdvisor by default #6314
• Rename ToolCallAdvisor to ToolCallingAdvisor with backward-compat shims and bug fixes #6303
• Fix OpenAiChatModel.stream() to only buffer tool calls. #6288
• refactor(chat-client): enforce non-null ToolCallAdvisor.Builder and fix advisor chain mutation #6282
• fix: fix span hierarchy in streaming paths for all remaining chat models #6276
• fix: bring parent span into ThreadLocals to maintain Observation and span hierarchy parity #6249
• Guard JSON schema generation #6217
• Fix merge in openai options #6215
• deepseek V4 Function calling 400 Bad Request from POST #5963

📔 Documentation

• Fix "vector strore" typo in RAG diagram #6300
• Polish reference documentation #6234
• Improve Mistral AI Chat documentation for reasoning models and ReasoningEffort Javadoc #6191
• Reintroduce community support documentation for OCI and Azure CosmosDB #6125

🔨 Dependency Upgrades

• Upgrade to MCP SDK 2.0.0-RC1 #6287

❤️ Contributors

Thank you to all the contributors who worked on this release:

@Baqirrizvidev, @Dream95, @EvanYao826, @ThomasVitale, @YunKuiLu, @chemicL, @dafriz, @ericbottard, @galenzo17, @goutamadwant, @guanxuc, @ilayaperumalg, @juanjuandog, @nicolaskrier, @quaff, @rajadilipkolli, @sdeleuze, @sobychacko, @stohirov, @symphony-enrico, @tzolov, and @wolf8334