Migrate Google Vertex AI integration from google-cloud-vertexai to google-genai (Imagen & Gemini)

Proposal: Migrate Google Vertex AI integration from google-cloud-vertexai to google-genai (Imagen & Gemini)

Summary

• Current state: The google-cloud-vertexai dependency is no longer the recommended path for Generative AI. In practice it’s convenient for Gemini chat flows but does not provide a first-class, maintainable path for Google’s Imagen image generation in Spring AI.
• Issue: With google-cloud-vertexai, Spring AI can wire up Gemini conversations, but cannot cleanly support Google’s Imagen image models via a portable ImageModel implementation.
• Proposal: Migrate the Google provider in Spring AI to Google Gen AI SDK (com.google.genai:google-genai) as the primary integration for both Gemini (chat/multimodal) and Imagen (image generation).

Reference repo for the current Vertex AI library:
https://github.com/googleapis/google-cloud-java/tree/main/java-vertexai

Background & Rationale

• Google recommends using the Google Gen AI SDK going forward.
• The google-genai SDK offers a unified Client that can target Gemini API or Vertex AI backends and includes built-in, high-level support for Imagen.
• Aligning Spring AI with google-genai improves feature coverage (Imagen), reduces low-level request handling, and fits Spring AI’s ImageModel abstraction.

Problem Statement

• The existing Spring AI adapter based on google-cloud-vertexai is oriented to Gemini chat and relies on low-level PredictionServiceClient#predict to hit image endpoints. This leads to ad-hoc request construction and an inconsistent developer experience versus other Spring AI models.
• We need a portable, first-class image generation flow for Google models (Imagen) in Spring AI.

Proposed Change

1. Introduce google-genai implementations
   • Add ImagenImageApi, ImagenImageOptions, and ImagenImageModel powered by com.google.genai:google-genai.
   • Expose Imagen parameters such as imageSize, aspectRatio, watermarking, language, safety filtering, and output formats through Spring AI’s options.
2. Update Gemini support
   • Provide a google-genai-based Gemini path for parity and consistency.
3. Documentation & samples
   • Document the migration and add end-to-end samples (Imagen generation with ImagePrompt → ImageResponse).

Compatibility & Migration

• Low-risk migration path: Keep the current google-cloud-vertexai adapter available for a transition period, but recommend google-genai for new development—especially for Imagen.
• Longer term: Prefer google-genai as the primary implementation and discourage adding new features on top of google-cloud-vertexai for image generation.

Alternatives Considered

• Continue on google-cloud-vertexai with custom wrappers around PredictionServiceClient#predict for Imagen.
  • Cons: Higher maintenance overhead, poorer developer experience, and not aligned with the recommended SDK direction.

References

• Vertex AI library repo (current): https://github.com/googleapis/google-cloud-java/tree/main/java-vertexai
• Google Gen AI SDK for Java (target): com.google.genai:google-genai

Tasks

• Add google-genai-based Imagen implementation (API/Options/Model).
• Update Gemini integration to use google-genai.
• Provide migration notes and samples.
• Add tests and documentation (Imagen generation, backend selection).

Start by inspecting the existing google-cloud-vertexai adapter and its use of PredictionServiceClient#predict. Then review the proposed ImagenImageApi, ImagenImageOptions, and ImagenImageModel entry points alongside the ImagePrompt → ImageResponse flow. Done means the google-genai Imagen path, Gemini support, migration notes, samples, tests, and backend-selection documentation are provided.