Serguey Asael Shinder: Spring AI 2.1.0-RC2 adds DecisionModel for typed yes, choice and score answers with probabilities
Spring AI 2.1.0-RC2 was announced on the Spring blog on 9 November by Dimitar Proynov and is available from Maven Central. It builds on the first release candidate and introduces one new model API, DecisionModel, together with the first features built on it. (https://spring.io/blog/2026/11/09/spring-ai-2-1-0-RC2-available-now)
The idea is simple. Many decisions in an application have a small, known set of results: yes or no, one of a few categories, or a level on a scale. Asking a chat model returns free text that the application must parse. A decision model instead picks one of the results you describe and reports how likely each one is. The API follows the shape of the other Spring AI model APIs: a DecisionPrompt holds an input and one or more named questions, a DecisionResponse comes back, and DecisionOptions tune the call. There are three question types: YesNoQuestion, answered with the probability of yes; ChoiceQuestion, answered with the selected option and, when the provider reports them, per-option probabilities; and ScoreQuestion, answered with a level of an ordered scale. Because answers carry probabilities, the application sets its own threshold, for example acting only when isYes(0.9) is true.

The post is careful to say that a decision model is still a machine learning model, not a rule engine, so its answers are not deterministic. Three implementations ship: OpenAiDecisionModel on the OpenAI Decisions API, OllamaDecisionModel, which runs locally with the d1-omni:600m model by default and needs no API key, and SystemOneDecisionModel on the TypeSafe Java SDK, which is not auto-configured.
Features already built on the API include DecisionDocumentReranker, which reorders retrieved documents by whether each answers the query, DecisionDocumentFilter, which screens documents for prompt injection, contradictions and relevance before they reach the prompt, and DecisionToolIndex, a fourth tool-search strategy next to the regex, Lucene and vector indexes. The reranker and filter fail open by default, keeping a document the model could not judge, and the filter can be set to fail closed. The release also adds tool call delta streaming, pre-computed embeddings for five more vector stores, server support for a new MCP protocol revision and an upgrade to the Anthropic Java SDK 3.x.
For a Spring application this is the first typed alternative to parsing chat output for routing and guardrails. As with any release candidate, it is the time to try the API on real code, and the fail-open default of the document filter is the setting to decide on deliberately.