Integrating Google Vertex AI: Expanding Model Choices in Dyad

Expanding Our AI Horizons

At Dyad, we have been continuously evolving how our application interacts with various AI providers. To offer more flexibility and model diversity, we have recently integrated Google Vertex AI as a first-class citizen alongside our existing support for OpenAI and local models via Ollama.

This integration allows users to leverage the powerful Gemini 2.5 series, including the Pro, Flash, and Flash-Lite models, directly through our platform by utilizing their own Google Cloud infrastructure.

Implementation Highlights

Seamless Provider Integration

By utilizing the official @ai-sdk/google-vertex package, we ensured that the new provider integrates naturally with our existing chat and streaming handlers. A key focus was providing a robust configuration UI where users can securely store their credentials.

// Example of initializing the Vertex provider
import { createVertex } from '@ai-sdk/google-vertex';

const vertex = createVertex({
  project: settings.projectId,
  location: settings.location,
  googleAuthOptions: {
    credentials: JSON.parse(settings.serviceAccountKey),
  },
});

This setup allows us to dynamically map built-in model IDs to the appropriate Vertex publisher paths, such as publishers/google/models/gemini-2.5-pro, ensuring that the resolution logic remains consistent regardless of the underlying infrastructure.

Controlled "Thinking" Capabilities

One of the most requested features was explicit control over AI reasoning behaviors. For the Gemini 2.5 Flash model, we implemented a toggle in our UI. This allows users to balance cost and reasoning depth based on their current task.

• Pro models: Always leverage deep thinking capabilities.
• Flash models: Users can toggle this behavior.
• Flash-Lite: Optimized for speed, reasoning is disabled by default.

Hardening Cross-Platform Operations

During development, we encountered inconsistencies in file path handling across different operating systems. By standardizing our internal file operations with toPosixPath and ensuring safeJoin semantics, we have guaranteed that the configuration files and model cache remain reliable whether running on macOS, Windows, or Linux within the Electron environment.

Verdict

Integrating Vertex AI has simplified the setup for our power users while providing a clear path to accessing cutting-edge models through Google Cloud.

Actionable Takeaway: If you are building a multi-provider AI application, prioritize a unified configuration schema. By treating credentials as encrypted settings and abstracting model resolution into a centralized service, you can add new providers without bloating your core business logic.

Generated with Gitvlg.com