kunthawat/moreminimore-vibe
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add Google Vertex AI provider (#1163)

Browse Source 
# Summary

* Adds first-class **Google Vertex AI provider** using
`@ai-sdk/google-vertex`.
* Supports **Gemini 2.5** models and partner **MaaS (Model Garden)**
models via full publisher IDs.
* New **Vertex-specific settings UI** for Project, Location, and Service
Account JSON.
* Implements a **“thinking” toggle** for Gemini 2.5 Flash

  * Pro: always on
  * Flash: toggleable
  * Flash Lite: none
* Fixes *“AI not found”* for Vertex built-ins by mapping to
`publishers/google` paths.
* Hardens **cross-platform file ops** and ensures all tests pass.

---

# What’s New

### Vertex AI Provider

* Uses `@ai-sdk/google-vertex` with `googleAuthOptions.credentials` from
pasted Service Account JSON.
* Configurable **project** and **location**.
* Base URL → `/projects/{project}/locations/{location}`

  * Built-ins: `publishers/google/models/<id>`
  * Partner MaaS: `publishers/<partner>/models/...`

### Built-in Vertex Models

* `gemini-2.5-pro`
* `gemini-2.5-flash`
* `gemini-2.5-flash-lite`

### Thinking Behavior

* Vertex + Google marked as thinking-capable.
* Pro: always thinking
* Flash: toggle in UI
* Flash Lite: none

### Vertex Settings UI

* New **Google Vertex AI panel** for Project ID, Location, Service
Account JSON.
* Keys encrypted like other secrets.

---

# Fixes

* **Model resolution:** built-ins auto-map to
`publishers/google/models/<id>`.
* **Partner MaaS support:** full publisher IDs work directly (e.g.
DeepSeek).
* **Cross-platform paths:** normalize file ops with `toPosixPath`,
preserve `safeJoin` semantics.

---

# Why This Is Better

* Users can select **Vertex alongside other providers**.
* **More models** available through Model Garden.
* **Dedicated setup UI** reduces misconfig.
* **Thinking toggle** gives control over cost vs. reasoning depth.

---

# Files Changed

* **Provider & Models**: `language_model_helpers.ts`,
`get_model_client.ts`
* **Streaming**: `chat_stream_handlers.ts`
* **Schemas & Encryption**: `schemas.ts`, `settings.ts`
* **Settings UI**: `VertexConfiguration.tsx`, `ApiKeyConfiguration.tsx`
* **Models UI**: `ModelsSection.tsx` (Flash toggle)
* **Setup Detection**: `useLanguageModelProviders.ts`
* **Path Utils**: `path_utils.ts`, `response_processor.ts`
* **Deps**: `package.json` → `@ai-sdk/google-vertex@3.0.16`

---

# Tests & Validation

* **TypeScript**: `npm run ts` → ✅
* **Lint**: `npm run lint` → ✅
* **Unit tests**: `npm test` → ✅ 231 passed, 0 failed

---

# Migration / Notes

* No breaking changes.
* For Vertex usage:

  * Ensure Vertex AI API is enabled.
  * Service Account needs `roles/aiplatform.user`.
  * Region must support model (e.g. `us-central1`).
* Thinking toggle currently affects **only** Gemini 2.5 Flash.

---

# Manual QA

1. Configure Vertex with Project/Location/Service Account JSON.
2. Test built-ins:

   * `gemini-2.5-pro`
   * `gemini-2.5-flash` (toggle on/off)
   * `gemini-2.5-flash-lite`
3. Test MaaS partner model (e.g., DeepSeek) via full publisher ID.
4. Verify other providers remain unaffected.
    
<!-- This is an auto-generated description by cubic. -->
---

## Summary by cubic
Adds a first-class Google Vertex AI provider with Gemini 2.5 models, a
Vertex settings panel, and a “thinking” toggle for Gemini 2.5 Flash.
Also fixes model resolution for Vertex and hardens cross-platform file
operations.

- **New Features**
- Vertex AI provider via @ai-sdk/google-vertex with project, location,
and service account JSON.
- Built-in models: gemini-2.5-pro, gemini-2.5-flash,
gemini-2.5-flash-lite.
- “Thinking” support: Pro always on; Flash toggle in Models UI; Flash
Lite none.
- MaaS partners supported via full publisher paths (e.g.,
publishers/<partner>/models/...).
  - Vertex settings UI with encrypted service account key storage.

- **Bug Fixes**
  - Built-in Vertex models auto-map to publishers/google/models/<id>.
  - Consistent file ops across platforms using toPosixPath.
- Vertex readiness detection requires project/location/service account
JSON.
- Streaming “thinking” behavior respects Vertex Flash toggle and Pro
always-on.

<!-- End of auto-generated description by cubic. -->

---------

Co-authored-by: Md Rakibul Islam Rocky <mdrirocky08@gmail.com>
Co-authored-by: graphite-app[bot] <96075541+graphite-app[bot]@users.noreply.github.com>
Co-authored-by: Will Chen <willchen90@gmail.com>
This commit is contained in:
Md Rakibul Islam Rocky
2025-09-09 08:41:12 +03:00
committed by GitHub
parent e962964afe
commit 4db6d63b72
11 changed files with 588 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
src/ipc/handlers/chat_stream_handlers.ts
View File

@@ -667,28 +667,53 @@ This conversation includes one or more image attachments. When the user uploads
} else {
logger.log("sending AI request");
}
// Build provider options with correct Google/Vertex thinking config gating
const providerOptions: Record<string, any> = {
"dyad-engine": {
dyadRequestId,
},
"dyad-gateway": getExtraProviderOptions(
modelClient.builtinProviderId,
settings,
),
openai: {
reasoningSummary: "auto",
} satisfies OpenAIResponsesProviderOptions,
};
// Conditionally include Google thinking config only for supported models
const selectedModelName = settings.selectedModel.name || "";
const providerId = modelClient.builtinProviderId;
const isVertex = providerId === "vertex";
const isGoogle = providerId === "google";
const isPartnerModel = selectedModelName.includes("/");
const isGeminiModel = selectedModelName.startsWith("gemini");
const isFlashLite = selectedModelName.includes("flash-lite");
// Keep Google provider behavior unchanged: always include includeThoughts
if (isGoogle) {
providerOptions.google = {
thinkingConfig: {
includeThoughts: true,
},
} satisfies GoogleGenerativeAIProviderOptions;
}
// Vertex-specific fix: only enable thinking on supported Gemini models
if (isVertex && isGeminiModel && !isFlashLite && !isPartnerModel) {
providerOptions.google = {
thinkingConfig: {
includeThoughts: true,
},
} satisfies GoogleGenerativeAIProviderOptions;
}
return streamText({
maxOutputTokens: await getMaxTokens(settings.selectedModel),
temperature: await getTemperature(settings.selectedModel),
maxRetries: 2,
model: modelClient.model,
providerOptions: {
"dyad-engine": {
dyadRequestId,
},
"dyad-gateway": getExtraProviderOptions(
modelClient.builtinProviderId,
settings,
),
google: {
thinkingConfig: {
includeThoughts: true,
},
} satisfies GoogleGenerativeAIProviderOptions,
openai: {
reasoningSummary: "auto",
} satisfies OpenAIResponsesProviderOptions,
},
providerOptions,
system: systemPrompt,
messages: chatMessages.filter((m) => m.content),
onError: (error: any) => {

29
src/ipc/shared/language_model_helpers.ts
View File

@@ -8,6 +8,7 @@ import { eq } from "drizzle-orm";
export const PROVIDERS_THAT_SUPPORT_THINKING: (keyof typeof MODEL_OPTIONS)[] = [
"google",
"vertex",
"auto",
];
@@ -144,6 +145,26 @@ export const MODEL_OPTIONS: Record<string, ModelOption[]> = {
dollarSigns: 2,
},
],
vertex: [
// Vertex Gemini 2.5 Pro
{
name: "gemini-2.5-pro",
displayName: "Gemini 2.5 Pro",
description: "Vertex Gemini 2.5 Pro",
maxOutputTokens: 65_536 - 1,
contextWindow: 1_048_576,
temperature: 0,
},
// Vertex Gemini 2.5 Flash
{
name: "gemini-2.5-flash",
displayName: "Gemini 2.5 Flash",
description: "Vertex Gemini 2.5 Flash",
maxOutputTokens: 65_536 - 1,
contextWindow: 1_048_576,
temperature: 0,
},
],
openrouter: [
{
name: "qwen/qwen3-coder",
@@ -270,6 +291,14 @@ export const CLOUD_PROVIDERS: Record<
websiteUrl: "https://aistudio.google.com/app/apikey",
gatewayPrefix: "gemini/",
},
vertex: {
displayName: "Google Vertex AI",
hasFreeTier: false,
websiteUrl: "https://console.cloud.google.com/vertex-ai",
// Use the same gateway prefix as Google Gemini for Dyad Pro compatibility.
gatewayPrefix: "gemini/",
secondary: true,
},
openrouter: {
displayName: "OpenRouter",
hasFreeTier: true,

46
src/ipc/utils/get_model_client.ts
View File

@@ -1,10 +1,15 @@
import { createOpenAI } from "@ai-sdk/openai";
import { createGoogleGenerativeAI as createGoogle } from "@ai-sdk/google";
import { createAnthropic } from "@ai-sdk/anthropic";
import { createVertex as createGoogleVertex } from "@ai-sdk/google-vertex";
import { azure } from "@ai-sdk/azure";
import { createOpenRouter } from "@openrouter/ai-sdk-provider";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import type { LargeLanguageModel, UserSettings } from "../../lib/schemas";
import type {
LargeLanguageModel,
UserSettings,
VertexProviderSetting,
} from "../../lib/schemas";
import { getEnvVar } from "./read_env";
import log from "electron-log";
import { getLanguageModelProviders } from "../shared/language_model_helpers";
@@ -216,6 +221,45 @@ function getRegularModelClient(
backupModelClients: [],
};
}
case "vertex": {
// Vertex uses Google service account credentials with project/location
const vertexSettings = settings.providerSettings?.[
model.provider
] as VertexProviderSetting;
const project = vertexSettings?.projectId;
const location = vertexSettings?.location;
const serviceAccountKey = vertexSettings?.serviceAccountKey?.value;
// Use a baseURL that does NOT pin to publishers/google so that
// full publisher model IDs (e.g. publishers/deepseek-ai/models/...) work.
const regionHost = `${location === "global" ? "" : `${location}-`}aiplatform.googleapis.com`;
const baseURL = `https://${regionHost}/v1/projects/${project}/locations/${location}`;
const provider = createGoogleVertex({
project,
location,
baseURL,
googleAuthOptions: serviceAccountKey
? {
// Expecting the user to paste the full JSON of the service account key
credentials: JSON.parse(serviceAccountKey),
}
: undefined,
});
return {
modelClient: {
// For built-in Google models on Vertex, the path must include
// publishers/google/models/<model>. For partner MaaS models the
// full publisher path is already included.
model: provider(
model.name.includes("/")
? model.name
: `publishers/google/models/${model.name}`,
),
builtinProviderId: providerId,
},
backupModelClients: [],
};
}
case "openrouter": {
const provider = createOpenRouter({ apiKey });
return {