@radaros/core 0.2.0 → 0.3.0

package/dist/index.d.ts

@@ -491,6 +491,11 @@ declare function google(modelId: string, config?: {
 declare function ollama(modelId: string, config?: {
     host?: string;
 }): ModelProvider;
+declare function vertex(modelId: string, config?: {
+    project?: string;
+    location?: string;
+    credentials?: string;
+}): ModelProvider;
 
 interface OpenAIConfig {
     apiKey?: string;
@@ -584,6 +589,42 @@ declare class OllamaProvider implements ModelProvider {
     private normalizeResponse;
 }
 
+interface VertexAIConfig {
+    project?: string;
+    location?: string;
+    /** Service account key JSON string or path (optional — uses ADC by default). */
+    credentials?: string;
+}
+/**
+ * Vertex AI provider using Google's @google/genai SDK in Vertex mode.
+ *
+ * Authentication (in order of precedence):
+ *   1. Explicit `project` + `location` in config
+ *   2. GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION env vars
+ *   3. Application Default Credentials (gcloud auth)
+ */
+declare class VertexAIProvider implements ModelProvider {
+    readonly providerId = "vertex";
+    readonly modelId: string;
+    private ai;
+    private GoogleGenAICtor;
+    private project;
+    private location;
+    constructor(modelId: string, config?: VertexAIConfig);
+    private getClient;
+    generate(messages: ChatMessage[], options?: ModelConfig & {
+        tools?: ToolDefinition[];
+    }): Promise<ModelResponse>;
+    stream(messages: ChatMessage[], options?: ModelConfig & {
+        tools?: ToolDefinition[];
+    }): AsyncGenerator<StreamChunk>;
+    private toGoogleMessages;
+    private partToGoogle;
+    private toGoogleTools;
+    private cleanJsonSchema;
+    private normalizeResponse;
+}
+
 declare function defineTool<T extends z.ZodObject<any>>(config: {
     name: string;
     description: string;
@@ -1107,4 +1148,4 @@ declare class A2ARemoteAgent {
     private partsToText;
 }
 
-export { type A2AAgentCard, type A2AArtifact, type A2ADataPart, type A2AFilePart, type A2AJsonRpcRequest, type A2AJsonRpcResponse, type A2AMessage, type A2APart, A2ARemoteAgent, type A2ARemoteAgentConfig, type A2ASendParams, type A2ASkill, type A2ATask, type A2ATaskQueryParams, type A2ATaskState, type A2ATextPart, Agent, type AgentConfig, type AgentEventMap, type AgentHooks, type AgentStep, AnthropicProvider, type Artifact, type AudioPart, BaseVectorStore, type ChatMessage, type ConditionStep, type ContentPart, type EmbeddingProvider, EventBus, type FilePart, type FunctionStep, GoogleEmbedding, type GoogleEmbeddingConfig, GoogleProvider, type GuardrailResult, type ImagePart, InMemoryStorage, InMemoryVectorStore, type InputGuardrail, KnowledgeBase, type KnowledgeBaseConfig, type KnowledgeBaseToolConfig, LLMLoop, type LogLevel, Logger, type LoggerConfig, MCPToolProvider, type MCPToolProviderConfig, Memory, type MemoryConfig, type MemoryEntry, type MessageContent, type MessageRole, type ModelConfig, type ModelProvider, ModelRegistry, type ModelResponse, MongoDBStorage, type MongoDBVectorConfig, MongoDBVectorStore, OllamaProvider, OpenAIEmbedding, type OpenAIEmbeddingConfig, OpenAIProvider, type OutputGuardrail, type ParallelStep, type PgVectorConfig, PgVectorStore, PostgresStorage, type QdrantConfig, QdrantVectorStore, RunContext, type RunOpts, type RunOutput, type Session, SessionManager, SqliteStorage, type StepDef, type StepResult, type StorageDriver, type StreamChunk, Team, type TeamConfig, TeamMode, type TextPart, type TokenUsage, type ToolCall, type ToolCallResult, type ToolDef, type ToolDefinition, ToolExecutor, type ToolResult, type VectorDocument, type VectorSearchOptions, type VectorSearchResult, type VectorStore, Workflow, type WorkflowConfig, type WorkflowResult, anthropic, defineTool, getTextContent, google, isMultiModal, ollama, openai, registry };
+export { type A2AAgentCard, type A2AArtifact, type A2ADataPart, type A2AFilePart, type A2AJsonRpcRequest, type A2AJsonRpcResponse, type A2AMessage, type A2APart, A2ARemoteAgent, type A2ARemoteAgentConfig, type A2ASendParams, type A2ASkill, type A2ATask, type A2ATaskQueryParams, type A2ATaskState, type A2ATextPart, Agent, type AgentConfig, type AgentEventMap, type AgentHooks, type AgentStep, AnthropicProvider, type Artifact, type AudioPart, BaseVectorStore, type ChatMessage, type ConditionStep, type ContentPart, type EmbeddingProvider, EventBus, type FilePart, type FunctionStep, GoogleEmbedding, type GoogleEmbeddingConfig, GoogleProvider, type GuardrailResult, type ImagePart, InMemoryStorage, InMemoryVectorStore, type InputGuardrail, KnowledgeBase, type KnowledgeBaseConfig, type KnowledgeBaseToolConfig, LLMLoop, type LogLevel, Logger, type LoggerConfig, MCPToolProvider, type MCPToolProviderConfig, Memory, type MemoryConfig, type MemoryEntry, type MessageContent, type MessageRole, type ModelConfig, type ModelProvider, ModelRegistry, type ModelResponse, MongoDBStorage, type MongoDBVectorConfig, MongoDBVectorStore, OllamaProvider, OpenAIEmbedding, type OpenAIEmbeddingConfig, OpenAIProvider, type OutputGuardrail, type ParallelStep, type PgVectorConfig, PgVectorStore, PostgresStorage, type QdrantConfig, QdrantVectorStore, RunContext, type RunOpts, type RunOutput, type Session, SessionManager, SqliteStorage, type StepDef, type StepResult, type StorageDriver, type StreamChunk, Team, type TeamConfig, TeamMode, type TextPart, type TokenUsage, type ToolCall, type ToolCallResult, type ToolDef, type ToolDefinition, ToolExecutor, type ToolResult, type VectorDocument, type VectorSearchOptions, type VectorSearchResult, type VectorStore, type VertexAIConfig, VertexAIProvider, Workflow, type WorkflowConfig, type WorkflowResult, anthropic, defineTool, getTextContent, google, isMultiModal, ollama, openai, registry, vertex };

package/dist/index.js

@@ -2367,6 +2367,291 @@ var OllamaProvider = class {
   }
 };
 
+// src/models/providers/vertex.ts
+import { createRequire as createRequire7 } from "module";
+var _require7 = createRequire7(import.meta.url);
+var VertexAIProvider = class {
+  providerId = "vertex";
+  modelId;
+  ai = null;
+  GoogleGenAICtor;
+  project;
+  location;
+  constructor(modelId, config) {
+    this.modelId = modelId;
+    this.project = config?.project ?? process.env.GOOGLE_CLOUD_PROJECT ?? process.env.GCLOUD_PROJECT ?? "";
+    this.location = config?.location ?? process.env.GOOGLE_CLOUD_LOCATION ?? process.env.GOOGLE_CLOUD_REGION ?? "us-central1";
+    if (!this.project) {
+      throw new Error(
+        "VertexAIProvider: 'project' is required. Pass it in config or set GOOGLE_CLOUD_PROJECT env var."
+      );
+    }
+    try {
+      const { GoogleGenAI } = _require7("@google/genai");
+      this.GoogleGenAICtor = GoogleGenAI;
+    } catch {
+      throw new Error(
+        "@google/genai is required for VertexAIProvider. Install it: npm install @google/genai"
+      );
+    }
+  }
+  getClient() {
+    if (this.ai) return this.ai;
+    this.ai = new this.GoogleGenAICtor({
+      vertexai: true,
+      project: this.project,
+      location: this.location
+    });
+    return this.ai;
+  }
+  async generate(messages, options) {
+    const { systemInstruction, contents } = this.toGoogleMessages(messages);
+    const config = {};
+    if (options?.temperature !== void 0)
+      config.temperature = options.temperature;
+    if (options?.maxTokens !== void 0)
+      config.maxOutputTokens = options.maxTokens;
+    if (options?.topP !== void 0) config.topP = options.topP;
+    if (options?.stop) config.stopSequences = options.stop;
+    if (options?.responseFormat) {
+      config.responseMimeType = "application/json";
+      const rf = options.responseFormat;
+      if (typeof rf === "object" && rf !== null && "type" in rf && rf.type === "json_schema" && "schema" in rf && rf.schema) {
+        config.responseSchema = this.cleanJsonSchema(
+          rf.schema
+        );
+      }
+    }
+    const params = {
+      model: this.modelId,
+      contents,
+      config
+    };
+    if (systemInstruction) params.systemInstruction = systemInstruction;
+    if (options?.tools?.length) {
+      params.tools = [
+        { functionDeclarations: this.toGoogleTools(options.tools) }
+      ];
+    }
+    const client = this.getClient();
+    const response = await client.models.generateContent(params);
+    return this.normalizeResponse(response);
+  }
+  async *stream(messages, options) {
+    const { systemInstruction, contents } = this.toGoogleMessages(messages);
+    const config = {};
+    if (options?.temperature !== void 0)
+      config.temperature = options.temperature;
+    if (options?.maxTokens !== void 0)
+      config.maxOutputTokens = options.maxTokens;
+    if (options?.topP !== void 0) config.topP = options.topP;
+    if (options?.stop) config.stopSequences = options.stop;
+    const params = {
+      model: this.modelId,
+      contents,
+      config
+    };
+    if (systemInstruction) params.systemInstruction = systemInstruction;
+    if (options?.tools?.length) {
+      params.tools = [
+        { functionDeclarations: this.toGoogleTools(options.tools) }
+      ];
+    }
+    const client = this.getClient();
+    const streamResult = await client.models.generateContentStream(params);
+    let toolCallCounter = 0;
+    for await (const chunk of streamResult) {
+      const candidate = chunk.candidates?.[0];
+      if (!candidate?.content?.parts) continue;
+      for (const part of candidate.content.parts) {
+        if (part.text) {
+          yield { type: "text", text: part.text };
+        }
+        if (part.functionCall) {
+          const id = `vertex_tc_${toolCallCounter++}`;
+          yield {
+            type: "tool_call_start",
+            toolCall: { id, name: part.functionCall.name }
+          };
+          yield {
+            type: "tool_call_delta",
+            toolCallId: id,
+            argumentsDelta: JSON.stringify(part.functionCall.args ?? {})
+          };
+          yield { type: "tool_call_end", toolCallId: id };
+        }
+      }
+      if (candidate.finishReason) {
+        let finishReason = "stop";
+        if (candidate.finishReason === "STOP" || candidate.finishReason === "END_TURN")
+          finishReason = "stop";
+        else if (candidate.finishReason === "MAX_TOKENS")
+          finishReason = "length";
+        else if (candidate.finishReason === "SAFETY")
+          finishReason = "content_filter";
+        const hasToolCalls = candidate.content?.parts?.some(
+          (p) => p.functionCall
+        );
+        if (hasToolCalls) finishReason = "tool_calls";
+        yield {
+          type: "finish",
+          finishReason,
+          usage: chunk.usageMetadata ? {
+            promptTokens: chunk.usageMetadata.promptTokenCount ?? 0,
+            completionTokens: chunk.usageMetadata.candidatesTokenCount ?? 0,
+            totalTokens: chunk.usageMetadata.totalTokenCount ?? 0
+          } : void 0
+        };
+      }
+    }
+  }
+  // ── Message conversion (identical to GoogleProvider) ─────────────────────
+  toGoogleMessages(messages) {
+    let systemInstruction;
+    const contents = [];
+    for (const msg of messages) {
+      if (msg.role === "system") {
+        systemInstruction = getTextContent(msg.content) || void 0;
+        continue;
+      }
+      if (msg.role === "user") {
+        if (isMultiModal(msg.content)) {
+          contents.push({
+            role: "user",
+            parts: msg.content.map((p) => this.partToGoogle(p))
+          });
+        } else {
+          contents.push({
+            role: "user",
+            parts: [{ text: msg.content ?? "" }]
+          });
+        }
+        continue;
+      }
+      if (msg.role === "assistant") {
+        const parts = [];
+        if (msg.content) parts.push({ text: msg.content });
+        if (msg.toolCalls) {
+          for (const tc of msg.toolCalls) {
+            parts.push({
+              functionCall: { name: tc.name, args: tc.arguments }
+            });
+          }
+        }
+        if (parts.length === 0) parts.push({ text: "" });
+        contents.push({ role: "model", parts });
+        continue;
+      }
+      if (msg.role === "tool") {
+        contents.push({
+          role: "function",
+          parts: [
+            {
+              functionResponse: {
+                name: msg.name ?? "unknown",
+                response: { result: msg.content ?? "" }
+              }
+            }
+          ]
+        });
+        continue;
+      }
+    }
+    return { systemInstruction, contents };
+  }
+  partToGoogle(part) {
+    switch (part.type) {
+      case "text":
+        return { text: part.text };
+      case "image":
+      case "audio":
+      case "file": {
+        const isUrl = part.data.startsWith("http://") || part.data.startsWith("https://");
+        if (isUrl) {
+          return {
+            fileData: {
+              fileUri: part.data,
+              mimeType: part.mimeType ?? (part.type === "image" ? "image/png" : "application/octet-stream")
+            }
+          };
+        }
+        return {
+          inlineData: {
+            data: part.data,
+            mimeType: part.mimeType ?? (part.type === "image" ? "image/png" : part.type === "audio" ? "audio/mp3" : "application/octet-stream")
+          }
+        };
+      }
+    }
+  }
+  toGoogleTools(tools) {
+    return tools.map((t) => ({
+      name: t.name,
+      description: t.description,
+      parameters: t.parameters
+    }));
+  }
+  cleanJsonSchema(schema) {
+    const cleaned = { ...schema };
+    delete cleaned["$schema"];
+    delete cleaned["$ref"];
+    delete cleaned["additionalProperties"];
+    if (cleaned.properties && typeof cleaned.properties === "object") {
+      const props = {};
+      for (const [key, val] of Object.entries(
+        cleaned.properties
+      )) {
+        props[key] = typeof val === "object" && val ? this.cleanJsonSchema(val) : val;
+      }
+      cleaned.properties = props;
+    }
+    if (cleaned.items && typeof cleaned.items === "object") {
+      cleaned.items = this.cleanJsonSchema(
+        cleaned.items
+      );
+    }
+    return cleaned;
+  }
+  normalizeResponse(response) {
+    const candidate = response.candidates?.[0];
+    const parts = candidate?.content?.parts ?? [];
+    let textContent = "";
+    const toolCalls = [];
+    let toolCallCounter = 0;
+    for (const part of parts) {
+      if (part.text) textContent += part.text;
+      if (part.functionCall) {
+        toolCalls.push({
+          id: `vertex_tc_${toolCallCounter++}`,
+          name: part.functionCall.name,
+          arguments: part.functionCall.args ?? {}
+        });
+      }
+    }
+    const usage = {
+      promptTokens: response.usageMetadata?.promptTokenCount ?? 0,
+      completionTokens: response.usageMetadata?.candidatesTokenCount ?? 0,
+      totalTokens: response.usageMetadata?.totalTokenCount ?? 0
+    };
+    let finishReason = "stop";
+    if (toolCalls.length > 0) finishReason = "tool_calls";
+    else if (candidate?.finishReason === "MAX_TOKENS")
+      finishReason = "length";
+    else if (candidate?.finishReason === "SAFETY")
+      finishReason = "content_filter";
+    return {
+      message: {
+        role: "assistant",
+        content: textContent || null,
+        toolCalls: toolCalls.length > 0 ? toolCalls : void 0
+      },
+      usage,
+      finishReason,
+      raw: response
+    };
+  }
+};
+
 // src/models/registry.ts
 var ModelRegistry = class {
   factories = /* @__PURE__ */ new Map();
@@ -2415,6 +2700,16 @@ function google(modelId, config) {
 function ollama(modelId, config) {
   return registry.resolve("ollama", modelId, config);
 }
+registry.register(
+  "vertex",
+  (modelId, config) => new VertexAIProvider(
+    modelId,
+    config
+  )
+);
+function vertex(modelId, config) {
+  return registry.resolve("vertex", modelId, config);
+}
 
 // src/tools/define-tool.ts
 function defineTool(config) {
@@ -2427,13 +2722,13 @@ function defineTool(config) {
 }
 
 // src/storage/sqlite.ts
-import { createRequire as createRequire7 } from "module";
-var _require7 = createRequire7(import.meta.url);
+import { createRequire as createRequire8 } from "module";
+var _require8 = createRequire8(import.meta.url);
 var SqliteStorage = class {
   db;
   constructor(dbPath) {
     try {
-      const Database = _require7("better-sqlite3");
+      const Database = _require8("better-sqlite3");
       this.db = new Database(dbPath);
       this.db.pragma("journal_mode = WAL");
       this.db.exec(`
@@ -2482,13 +2777,13 @@ var SqliteStorage = class {
 };
 
 // src/storage/postgres.ts
-import { createRequire as createRequire8 } from "module";
-var _require8 = createRequire8(import.meta.url);
+import { createRequire as createRequire9 } from "module";
+var _require9 = createRequire9(import.meta.url);
 var PostgresStorage = class {
   pool;
   constructor(connectionString) {
     try {
-      const { Pool } = _require8("pg");
+      const { Pool } = _require9("pg");
       this.pool = new Pool({ connectionString });
     } catch {
       throw new Error(
@@ -2549,15 +2844,15 @@ var PostgresStorage = class {
 };
 
 // src/storage/mongodb.ts
-import { createRequire as createRequire9 } from "module";
-var _require9 = createRequire9(import.meta.url);
+import { createRequire as createRequire10 } from "module";
+var _require10 = createRequire10(import.meta.url);
 var MongoDBStorage = class {
   constructor(uri, dbName = "radaros", collectionName = "kv_store") {
     this.uri = uri;
     this.dbName = dbName;
     this.collectionName = collectionName;
     try {
-      const { MongoClient } = _require9("mongodb");
+      const { MongoClient } = _require10("mongodb");
       this.client = new MongoClient(uri);
     } catch {
       throw new Error(
@@ -2717,8 +3012,8 @@ var InMemoryVectorStore = class extends BaseVectorStore {
 };
 
 // src/vector/pgvector.ts
-import { createRequire as createRequire10 } from "module";
-var _require10 = createRequire10(import.meta.url);
+import { createRequire as createRequire11 } from "module";
+var _require11 = createRequire11(import.meta.url);
 var PgVectorStore = class extends BaseVectorStore {
   pool;
   dimensions;
@@ -2727,7 +3022,7 @@ var PgVectorStore = class extends BaseVectorStore {
     super(embedder);
     this.dimensions = config.dimensions ?? embedder?.dimensions ?? 1536;
     try {
-      const { Pool } = _require10("pg");
+      const { Pool } = _require11("pg");
       this.pool = new Pool({ connectionString: config.connectionString });
     } catch {
       throw new Error(
@@ -2846,9 +3141,9 @@ var PgVectorStore = class extends BaseVectorStore {
 };
 
 // src/vector/qdrant.ts
-import { createRequire as createRequire11 } from "module";
+import { createRequire as createRequire12 } from "module";
 import { createHash } from "crypto";
-var _require11 = createRequire11(import.meta.url);
+var _require12 = createRequire12(import.meta.url);
 var UUID_RE = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
 function stringToUUID(str) {
   const hex = createHash("md5").update(str).digest("hex");
@@ -2873,7 +3168,7 @@ var QdrantVectorStore = class extends BaseVectorStore {
     super(embedder);
     this.dimensions = config.dimensions ?? embedder?.dimensions ?? 1536;
     try {
-      const { QdrantClient } = _require11("@qdrant/js-client-rest");
+      const { QdrantClient } = _require12("@qdrant/js-client-rest");
       this.client = new QdrantClient({
         url: config.url ?? "http://localhost:6333",
         apiKey: config.apiKey,
@@ -3006,8 +3301,8 @@ var QdrantVectorStore = class extends BaseVectorStore {
 };
 
 // src/vector/mongodb.ts
-import { createRequire as createRequire12 } from "module";
-var _require12 = createRequire12(import.meta.url);
+import { createRequire as createRequire13 } from "module";
+var _require13 = createRequire13(import.meta.url);
 var MongoDBVectorStore = class extends BaseVectorStore {
   client;
   db;
@@ -3019,7 +3314,7 @@ var MongoDBVectorStore = class extends BaseVectorStore {
     this.indexName = config.indexName ?? "vector_index";
     this.dbName = config.dbName ?? "radaros_vectors";
     try {
-      const { MongoClient } = _require12("mongodb");
+      const { MongoClient } = _require13("mongodb");
       this.client = new MongoClient(config.uri);
     } catch {
       throw new Error(
@@ -3190,8 +3485,8 @@ function cosine(a, b) {
 }
 
 // src/vector/embeddings/openai.ts
-import { createRequire as createRequire13 } from "module";
-var _require13 = createRequire13(import.meta.url);
+import { createRequire as createRequire14 } from "module";
+var _require14 = createRequire14(import.meta.url);
 var MODEL_DIMENSIONS = {
   "text-embedding-3-small": 1536,
   "text-embedding-3-large": 3072,
@@ -3205,7 +3500,7 @@ var OpenAIEmbedding = class {
     this.model = config.model ?? "text-embedding-3-small";
     this.dimensions = config.dimensions ?? MODEL_DIMENSIONS[this.model] ?? 1536;
     try {
-      const mod = _require13("openai");
+      const mod = _require14("openai");
       const OpenAI = mod.default ?? mod;
       this.client = new OpenAI({
         apiKey: config.apiKey ?? process.env.OPENAI_API_KEY,
@@ -3236,8 +3531,8 @@ var OpenAIEmbedding = class {
 };
 
 // src/vector/embeddings/google.ts
-import { createRequire as createRequire14 } from "module";
-var _require14 = createRequire14(import.meta.url);
+import { createRequire as createRequire15 } from "module";
+var _require15 = createRequire15(import.meta.url);
 var MODEL_DIMENSIONS2 = {
   "text-embedding-004": 768,
   "embedding-001": 768
@@ -3250,7 +3545,7 @@ var GoogleEmbedding = class {
     this.model = config.model ?? "text-embedding-004";
     this.dimensions = config.dimensions ?? MODEL_DIMENSIONS2[this.model] ?? 768;
     try {
-      const { GoogleGenAI } = _require14("@google/genai");
+      const { GoogleGenAI } = _require15("@google/genai");
       this.ai = new GoogleGenAI({
         apiKey: config.apiKey ?? process.env.GOOGLE_API_KEY
       });
@@ -3843,6 +4138,7 @@ export {
   Team,
   TeamMode,
   ToolExecutor,
+  VertexAIProvider,
   Workflow,
   anthropic,
   defineTool,
@@ -3851,5 +4147,6 @@ export {
   isMultiModal,
   ollama,
   openai,
-  registry
+  registry,
+  vertex
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@radaros/core",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",

package/src/index.ts

@@ -49,11 +49,13 @@ export type {
   ModelConfig,
 } from "./models/types.js";
 export { getTextContent, isMultiModal } from "./models/types.js";
-export { ModelRegistry, registry, openai, anthropic, google, ollama } from "./models/registry.js";
+export { ModelRegistry, registry, openai, anthropic, google, ollama, vertex } from "./models/registry.js";
 export { OpenAIProvider } from "./models/providers/openai.js";
 export { AnthropicProvider } from "./models/providers/anthropic.js";
 export { GoogleProvider } from "./models/providers/google.js";
 export { OllamaProvider } from "./models/providers/ollama.js";
+export { VertexAIProvider } from "./models/providers/vertex.js";
+export type { VertexAIConfig } from "./models/providers/vertex.js";
 
 // Tools
 export { defineTool } from "./tools/define-tool.js";

package/src/models/providers/vertex.ts

@@ -0,0 +1,400 @@
+import { createRequire } from "node:module";
+import type { ModelProvider } from "../provider.js";
+import {
+  getTextContent,
+  isMultiModal,
+  type ChatMessage,
+  type ContentPart,
+  type ModelConfig,
+  type ModelResponse,
+  type StreamChunk,
+  type ToolDefinition,
+  type TokenUsage,
+  type ToolCall,
+} from "../types.js";
+
+const _require = createRequire(import.meta.url);
+
+export interface VertexAIConfig {
+  project?: string;
+  location?: string;
+  /** Service account key JSON string or path (optional — uses ADC by default). */
+  credentials?: string;
+}
+
+/**
+ * Vertex AI provider using Google's @google/genai SDK in Vertex mode.
+ *
+ * Authentication (in order of precedence):
+ *   1. Explicit `project` + `location` in config
+ *   2. GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION env vars
+ *   3. Application Default Credentials (gcloud auth)
+ */
+export class VertexAIProvider implements ModelProvider {
+  readonly providerId = "vertex";
+  readonly modelId: string;
+  private ai: any = null;
+  private GoogleGenAICtor: any;
+  private project: string;
+  private location: string;
+
+  constructor(modelId: string, config?: VertexAIConfig) {
+    this.modelId = modelId;
+    this.project =
+      config?.project ??
+      process.env.GOOGLE_CLOUD_PROJECT ??
+      process.env.GCLOUD_PROJECT ??
+      "";
+    this.location =
+      config?.location ??
+      process.env.GOOGLE_CLOUD_LOCATION ??
+      process.env.GOOGLE_CLOUD_REGION ??
+      "us-central1";
+
+    if (!this.project) {
+      throw new Error(
+        "VertexAIProvider: 'project' is required. Pass it in config or set GOOGLE_CLOUD_PROJECT env var."
+      );
+    }
+
+    try {
+      const { GoogleGenAI } = _require("@google/genai");
+      this.GoogleGenAICtor = GoogleGenAI;
+    } catch {
+      throw new Error(
+        "@google/genai is required for VertexAIProvider. Install it: npm install @google/genai"
+      );
+    }
+  }
+
+  private getClient(): any {
+    if (this.ai) return this.ai;
+    this.ai = new this.GoogleGenAICtor({
+      vertexai: true,
+      project: this.project,
+      location: this.location,
+    });
+    return this.ai;
+  }
+
+  async generate(
+    messages: ChatMessage[],
+    options?: ModelConfig & { tools?: ToolDefinition[] }
+  ): Promise<ModelResponse> {
+    const { systemInstruction, contents } = this.toGoogleMessages(messages);
+
+    const config: Record<string, unknown> = {};
+    if (options?.temperature !== undefined)
+      config.temperature = options.temperature;
+    if (options?.maxTokens !== undefined)
+      config.maxOutputTokens = options.maxTokens;
+    if (options?.topP !== undefined) config.topP = options.topP;
+    if (options?.stop) config.stopSequences = options.stop;
+
+    if (options?.responseFormat) {
+      config.responseMimeType = "application/json";
+      const rf = options.responseFormat;
+      if (
+        typeof rf === "object" &&
+        rf !== null &&
+        "type" in rf &&
+        rf.type === "json_schema" &&
+        "schema" in rf &&
+        rf.schema
+      ) {
+        config.responseSchema = this.cleanJsonSchema(
+          rf.schema as Record<string, unknown>
+        );
+      }
+    }
+
+    const params: Record<string, unknown> = {
+      model: this.modelId,
+      contents,
+      config,
+    };
+
+    if (systemInstruction) params.systemInstruction = systemInstruction;
+    if (options?.tools?.length) {
+      params.tools = [
+        { functionDeclarations: this.toGoogleTools(options.tools) },
+      ];
+    }
+
+    const client = this.getClient();
+    const response = await client.models.generateContent(params);
+    return this.normalizeResponse(response);
+  }
+
+  async *stream(
+    messages: ChatMessage[],
+    options?: ModelConfig & { tools?: ToolDefinition[] }
+  ): AsyncGenerator<StreamChunk> {
+    const { systemInstruction, contents } = this.toGoogleMessages(messages);
+
+    const config: Record<string, unknown> = {};
+    if (options?.temperature !== undefined)
+      config.temperature = options.temperature;
+    if (options?.maxTokens !== undefined)
+      config.maxOutputTokens = options.maxTokens;
+    if (options?.topP !== undefined) config.topP = options.topP;
+    if (options?.stop) config.stopSequences = options.stop;
+
+    const params: Record<string, unknown> = {
+      model: this.modelId,
+      contents,
+      config,
+    };
+
+    if (systemInstruction) params.systemInstruction = systemInstruction;
+    if (options?.tools?.length) {
+      params.tools = [
+        { functionDeclarations: this.toGoogleTools(options.tools) },
+      ];
+    }
+
+    const client = this.getClient();
+    const streamResult = await client.models.generateContentStream(params);
+
+    let toolCallCounter = 0;
+
+    for await (const chunk of streamResult) {
+      const candidate = chunk.candidates?.[0];
+      if (!candidate?.content?.parts) continue;
+
+      for (const part of candidate.content.parts) {
+        if (part.text) {
+          yield { type: "text", text: part.text };
+        }
+
+        if (part.functionCall) {
+          const id = `vertex_tc_${toolCallCounter++}`;
+          yield {
+            type: "tool_call_start",
+            toolCall: { id, name: part.functionCall.name },
+          };
+          yield {
+            type: "tool_call_delta",
+            toolCallId: id,
+            argumentsDelta: JSON.stringify(part.functionCall.args ?? {}),
+          };
+          yield { type: "tool_call_end", toolCallId: id };
+        }
+      }
+
+      if (candidate.finishReason) {
+        let finishReason = "stop";
+        if (
+          candidate.finishReason === "STOP" ||
+          candidate.finishReason === "END_TURN"
+        )
+          finishReason = "stop";
+        else if (candidate.finishReason === "MAX_TOKENS")
+          finishReason = "length";
+        else if (candidate.finishReason === "SAFETY")
+          finishReason = "content_filter";
+
+        const hasToolCalls = candidate.content?.parts?.some(
+          (p: any) => p.functionCall
+        );
+        if (hasToolCalls) finishReason = "tool_calls";
+
+        yield {
+          type: "finish",
+          finishReason,
+          usage: chunk.usageMetadata
+            ? {
+                promptTokens: chunk.usageMetadata.promptTokenCount ?? 0,
+                completionTokens:
+                  chunk.usageMetadata.candidatesTokenCount ?? 0,
+                totalTokens: chunk.usageMetadata.totalTokenCount ?? 0,
+              }
+            : undefined,
+        };
+      }
+    }
+  }
+
+  // ── Message conversion (identical to GoogleProvider) ─────────────────────
+
+  private toGoogleMessages(messages: ChatMessage[]): {
+    systemInstruction: string | undefined;
+    contents: unknown[];
+  } {
+    let systemInstruction: string | undefined;
+    const contents: unknown[] = [];
+
+    for (const msg of messages) {
+      if (msg.role === "system") {
+        systemInstruction = getTextContent(msg.content) || undefined;
+        continue;
+      }
+
+      if (msg.role === "user") {
+        if (isMultiModal(msg.content)) {
+          contents.push({
+            role: "user",
+            parts: msg.content.map((p) => this.partToGoogle(p)),
+          });
+        } else {
+          contents.push({
+            role: "user",
+            parts: [{ text: msg.content ?? "" }],
+          });
+        }
+        continue;
+      }
+
+      if (msg.role === "assistant") {
+        const parts: unknown[] = [];
+        if (msg.content) parts.push({ text: msg.content });
+        if (msg.toolCalls) {
+          for (const tc of msg.toolCalls) {
+            parts.push({
+              functionCall: { name: tc.name, args: tc.arguments },
+            });
+          }
+        }
+        if (parts.length === 0) parts.push({ text: "" });
+        contents.push({ role: "model", parts });
+        continue;
+      }
+
+      if (msg.role === "tool") {
+        contents.push({
+          role: "function",
+          parts: [
+            {
+              functionResponse: {
+                name: msg.name ?? "unknown",
+                response: { result: msg.content ?? "" },
+              },
+            },
+          ],
+        });
+        continue;
+      }
+    }
+
+    return { systemInstruction, contents };
+  }
+
+  private partToGoogle(part: ContentPart): unknown {
+    switch (part.type) {
+      case "text":
+        return { text: part.text };
+      case "image":
+      case "audio":
+      case "file": {
+        const isUrl =
+          part.data.startsWith("http://") || part.data.startsWith("https://");
+        if (isUrl) {
+          return {
+            fileData: {
+              fileUri: part.data,
+              mimeType:
+                part.mimeType ??
+                (part.type === "image"
+                  ? "image/png"
+                  : "application/octet-stream"),
+            },
+          };
+        }
+        return {
+          inlineData: {
+            data: part.data,
+            mimeType:
+              part.mimeType ??
+              (part.type === "image"
+                ? "image/png"
+                : part.type === "audio"
+                  ? "audio/mp3"
+                  : "application/octet-stream"),
+          },
+        };
+      }
+    }
+  }
+
+  private toGoogleTools(tools: ToolDefinition[]): unknown[] {
+    return tools.map((t) => ({
+      name: t.name,
+      description: t.description,
+      parameters: t.parameters,
+    }));
+  }
+
+  private cleanJsonSchema(
+    schema: Record<string, unknown>
+  ): Record<string, unknown> {
+    const cleaned = { ...schema };
+    delete cleaned["$schema"];
+    delete cleaned["$ref"];
+    delete cleaned["additionalProperties"];
+
+    if (cleaned.properties && typeof cleaned.properties === "object") {
+      const props: Record<string, unknown> = {};
+      for (const [key, val] of Object.entries(
+        cleaned.properties as Record<string, unknown>
+      )) {
+        props[key] =
+          typeof val === "object" && val
+            ? this.cleanJsonSchema(val as Record<string, unknown>)
+            : val;
+      }
+      cleaned.properties = props;
+    }
+
+    if (cleaned.items && typeof cleaned.items === "object") {
+      cleaned.items = this.cleanJsonSchema(
+        cleaned.items as Record<string, unknown>
+      );
+    }
+
+    return cleaned;
+  }
+
+  private normalizeResponse(response: any): ModelResponse {
+    const candidate = response.candidates?.[0];
+    const parts = candidate?.content?.parts ?? [];
+
+    let textContent = "";
+    const toolCalls: ToolCall[] = [];
+    let toolCallCounter = 0;
+
+    for (const part of parts) {
+      if (part.text) textContent += part.text;
+      if (part.functionCall) {
+        toolCalls.push({
+          id: `vertex_tc_${toolCallCounter++}`,
+          name: part.functionCall.name,
+          arguments: part.functionCall.args ?? {},
+        });
+      }
+    }
+
+    const usage: TokenUsage = {
+      promptTokens: response.usageMetadata?.promptTokenCount ?? 0,
+      completionTokens: response.usageMetadata?.candidatesTokenCount ?? 0,
+      totalTokens: response.usageMetadata?.totalTokenCount ?? 0,
+    };
+
+    let finishReason: ModelResponse["finishReason"] = "stop";
+    if (toolCalls.length > 0) finishReason = "tool_calls";
+    else if (candidate?.finishReason === "MAX_TOKENS")
+      finishReason = "length";
+    else if (candidate?.finishReason === "SAFETY")
+      finishReason = "content_filter";
+
+    return {
+      message: {
+        role: "assistant",
+        content: textContent || null,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+      },
+      usage,
+      finishReason,
+      raw: response,
+    };
+  }
+}

package/src/models/registry.ts

@@ -3,6 +3,7 @@ import { OpenAIProvider } from "./providers/openai.js";
 import { AnthropicProvider } from "./providers/anthropic.js";
 import { GoogleProvider } from "./providers/google.js";
 import { OllamaProvider } from "./providers/ollama.js";
+import { VertexAIProvider } from "./providers/vertex.js";
 
 type ProviderFactory = (
   modelId: string,
@@ -88,3 +89,19 @@ export function ollama(
 ): ModelProvider {
   return registry.resolve("ollama", modelId, config);
 }
+
+registry.register(
+  "vertex",
+  (modelId, config) =>
+    new VertexAIProvider(
+      modelId,
+      config as { project?: string; location?: string; credentials?: string }
+    )
+);
+
+export function vertex(
+  modelId: string,
+  config?: { project?: string; location?: string; credentials?: string }
+): ModelProvider {
+  return registry.resolve("vertex", modelId, config);
+}