Spring AI 2.1.0-M1: OpenAI Responses API, ordered MessagePart multimodal, VectorStore.upsert for precomputed embeddings

Core Background & Industry Pain Points Enterprise Java stacks struggle to fold Responses-style reasoning signatures, tool calls, multimodal attachments, and offline precomputed embeddings into one Chat/RAG abstraction. Older Spring AI leaned on Chat Completions field bags—order gets lost—and VectorStore assumed on-the-fly EmbeddingModel calls, which blocks external feature pipelines. Responses also rejects unsigned reasoning replay without an opaque payload. ### Architecture Highlights & Internals v2.1.0-M1 ships three pieces. OpenAI Responses mapping turns items into ReasoningPart (OpaquePayload), TextPart, ToolCallPart, MediaPart, UnknownPart; only function_call becomes ToolCalls. MessagePart gives Assistant/User/ToolResponse an ordered shared list; legacy accessors remain views. VectorStore.upsert takes EmbeddedDocument pairs and replace-by-id on pgvector/Redis/Elasticsearch/Qdrant with UPSERT observation. BOM targets Spring Boot 4.2. ### Authoritative Benchmarks & Measured Scores No public cross-version latency or recall numbers in the notes. Validate Responses multi-turn replay (signed reasoning), MessagePart order vs provider items, and upsert round-trips. Do not extrapolate Completions latency to Responses. ### Developer Hands-on Guide Use spring-ai-bom:2.1.0-M1, read the 2.1 upgrade notes and getting started. Milestone APIs may still move—canary before GA.