generate-spring-ai-feature

name: generate-spring-ai-feature description: Generates a Spring AI 2.0.0-M6 feature on Spring Boot 4.0.6 — ChatClient calls, Advisors, RAG with PgVector, structured output, tool/function calling. Use when the user asks for an AI capability (chat, summarisation, classification, agentic tools).

generate-spring-ai-feature

When to use

Any backend feature that calls a chat model, embeds documents, runs RAG, or exposes tools to a model.

Required design choices

• API surface: a typed interface describing inputs/outputs as records. Never expose raw String.
• ChatClient: built once via ChatClient.Builder, configured with default Advisors.
• Advisors: at least logging + safety. Add MessageChatMemoryAdvisor if the feature is conversational. Add a RAG advisor if grounding is required.
• Structured output: target a record. Use Spring AI's structured output converters — never parse JSON by hand.
• Tool/function calling: declare a typed callback interface, annotate with @Tool from Spring AI 2.0. Always supply explicit conversationId to chat memory advisors.
• Vector store: PgVector if PostgreSQL is present, else the project's default. Configure dimension to match the chosen embedding model.
• Prompts: live in src/main/resources/prompts/<feature>.st, never inline.

Required rules

• No secrets in code. API keys via env vars + @ConfigurationProperties record.
• Every model input crossing a tool boundary is validated and sanitised to prevent prompt injection.
• Streaming endpoints use Flux<String> (server-sent events) or Flow if Kotlin.
• Invoke generate-junit-test and generate-integration-test (with Testcontainers Ollama or wiremock) after generation.

Validation

• validate-java25-idioms, enforce-documentation, JUnit suite green.
• scan-source-vulnerabilities finds no prompt-injection vector.

Failure modes

• No vector store configured but RAG requested → ask architect.
• Chosen embedding dimension differs from existing index → stop.