@contentgrowth/llm-service 1.1.2 → 1.2.1

package/dist/index.js

@@ -71,8 +71,10 @@ var DefaultConfigProvider = class extends BaseConfigProvider {
   _buildTenantConfig(tenantConfig, env) {
     return {
       provider: tenantConfig.provider,
-      apiKey: tenantConfig.api_key,
       models: MODEL_CONFIGS[tenantConfig.provider],
+      apiKey: tenantConfig.api_key,
+      project: tenantConfig.project,
+      location: tenantConfig.location,
       temperature: parseFloat(env.DEFAULT_TEMPERATURE || "0.7"),
       maxTokens: parseInt(env.DEFAULT_MAX_TOKENS || "65536"),
       capabilities: tenantConfig.capabilities || { chat: true, image: false, video: false },
@@ -105,6 +107,28 @@ var DefaultConfigProvider = class extends BaseConfigProvider {
         image: env.GEMINI_IMAGE_MODEL || providerDefaults.image,
         video: env.GEMINI_VIDEO_MODEL || providerDefaults.video
       };
+    } else if (provider === "vertex") {
+      apiKey = env.VERTEX_API_KEY;
+      const project = env.VERTEX_PROJECT || env.GOOGLE_CLOUD_PROJECT;
+      const location = env.VERTEX_LOCATION || env.GOOGLE_CLOUD_LOCATION || "us-central1";
+      models = {
+        default: env.VERTEX_MODEL || providerDefaults.default,
+        edge: env.VERTEX_MODEL_EDGE || providerDefaults.edge,
+        fast: env.VERTEX_MODEL_FAST || providerDefaults.fast,
+        cost: env.VERTEX_MODEL_COST || providerDefaults.cost,
+        free: env.VERTEX_MODEL_FREE || providerDefaults.free,
+        image: env.VERTEX_IMAGE_MODEL || providerDefaults.image,
+        video: env.VERTEX_VIDEO_MODEL || providerDefaults.video
+      };
+      return {
+        provider,
+        apiKey,
+        project,
+        location,
+        models,
+        temperature: parseFloat(env.DEFAULT_TEMPERATURE || "0.7"),
+        maxTokens: parseInt(env.DEFAULT_MAX_TOKENS || "65536")
+      };
     }
     return {
       provider,
@@ -139,6 +163,15 @@ var MODEL_CONFIGS = {
     video: "veo",
     image: "gemini-3-pro-image-preview"
     // Default image generation model
+  },
+  vertex: {
+    default: "gemini-3-flash-preview",
+    edge: "gemini-3-pro-preview",
+    fast: "gemini-3-flash-preview",
+    cost: "gemini-3-flash-preview",
+    free: "gemini-3-flash-preview",
+    video: "veo",
+    image: "gemini-3-pro-image-preview"
   }
 };
 var ConfigManager = class {
@@ -536,25 +569,43 @@ var OpenAIProvider = class extends BaseLLMProvider {
   }
 };
 
-// src/llm/providers/gemini-provider.js
+// src/llm/providers/google-provider.js
 import { GoogleGenAI } from "@google/genai";
-var GeminiProvider = class extends BaseLLMProvider {
+var GoogleProvider = class extends BaseLLMProvider {
   constructor(config) {
     super(config);
-    const clientConfig = {};
-    if (config.project || config.location) {
-      console.log(`[GeminiProvider] Initializing with Vertex AI (Project: ${config.project}, Location: ${config.location || "us-central1"})`);
-      clientConfig.vertexAI = {
-        project: config.project,
-        location: config.location || "us-central1"
-      };
-    } else {
-      clientConfig.apiKey = config.apiKey;
-    }
-    this.client = new GoogleGenAI(clientConfig);
     this.models = config.models;
     this.defaultModel = config.models.default;
     this._pendingOperations = /* @__PURE__ */ new Map();
+    if (config.provider === "vertex") {
+      if (config.apiKey) {
+        this.client = new GoogleGenAI({
+          vertexai: true,
+          apiKey: config.apiKey
+        });
+      } else {
+        if (!config.project) {
+          console.warn("[GoogleProvider] Vertex AI: no project ID and no API key. Calls will likely fail.");
+        }
+        this.client = new GoogleGenAI({
+          vertexai: true,
+          project: config.project,
+          location: config.location || "us-central1"
+        });
+      }
+    } else {
+      this.client = new GoogleGenAI({
+        apiKey: config.apiKey
+      });
+    }
+  }
+  /**
+   * Perform the actual API call. Both AI Studio and Vertex AI use the
+   * same @google/genai SDK method — the routing is determined by how
+   * the client was constructed.
+   */
+  async _generateContent(requestOptions) {
+    return this.client.models.generateContent(requestOptions);
   }
   async chat(userMessage, systemPrompt = "", options = {}) {
     const messages = [{ role: "user", content: userMessage }];
@@ -586,7 +637,7 @@ var GeminiProvider = class extends BaseLLMProvider {
     );
   }
   async _chatCompletionWithModel(messages, systemPrompt, tools, modelName, maxTokens, temperature, options = {}) {
-    var _a, _b, _c, _d, _e, _f, _g;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k;
     const generationConfig = {
       temperature: (_a = options.temperature) != null ? _a : temperature,
       maxOutputTokens: (_b = options.maxTokens) != null ? _b : maxTokens
@@ -612,8 +663,41 @@ ${msg.content}`;
         }
       }
     }
-    const contents = geminiMessages.map((msg, index) => {
-      var _a2, _b2, _c2, _d2;
+    const contents = [];
+    let pendingToolParts = [];
+    for (let index = 0; index < geminiMessages.length; index++) {
+      const msg = geminiMessages[index];
+      if (msg.role === "tool") {
+        let assistantMsg = null;
+        for (let j = index - 1; j >= 0; j--) {
+          if (geminiMessages[j].role === "assistant" && geminiMessages[j].tool_calls) {
+            assistantMsg = geminiMessages[j];
+            break;
+          }
+        }
+        const toolCall = (_c = assistantMsg == null ? void 0 : assistantMsg.tool_calls) == null ? void 0 : _c.find((tc) => tc.id === msg.tool_call_id);
+        pendingToolParts.push({
+          functionResponse: {
+            name: ((_d = toolCall == null ? void 0 : toolCall.function) == null ? void 0 : _d.name) || "unknown_tool",
+            response: { content: msg.content }
+          }
+        });
+        const nextMsg = geminiMessages[index + 1];
+        if (!nextMsg || nextMsg.role !== "tool") {
+          if (options.responseFormat === "json" || ((_e = options.responseFormat) == null ? void 0 : _e.type) === "json_schema" || options.responseSchema) {
+            pendingToolParts.push({ text: "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]" });
+          } else {
+            pendingToolParts.push({ text: "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]" });
+          }
+          contents.push({ role: "user", parts: pendingToolParts });
+          pendingToolParts = [];
+        }
+        continue;
+      }
+      if (pendingToolParts.length > 0) {
+        contents.push({ role: "user", parts: pendingToolParts });
+        pendingToolParts = [];
+      }
       let role = "";
       let parts2;
       switch (msg.role) {
@@ -622,7 +706,7 @@ ${msg.content}`;
           parts2 = [{ text: msg.content }];
           if (index === geminiMessages.length - 1) {
             let reminder = "";
-            if (options.responseFormat === "json" || ((_a2 = options.responseFormat) == null ? void 0 : _a2.type) === "json_schema" || options.responseSchema) {
+            if (options.responseFormat === "json" || ((_f = options.responseFormat) == null ? void 0 : _f.type) === "json_schema" || options.responseSchema) {
               reminder = "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]";
             } else {
               reminder = "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]";
@@ -656,27 +740,11 @@ ${msg.content}`;
             parts2 = [part];
           }
           break;
-        case "tool":
-          role = "user";
-          const preceding_message = messages[index - 1];
-          const tool_call = (_b2 = preceding_message == null ? void 0 : preceding_message.tool_calls) == null ? void 0 : _b2.find((tc) => tc.id === msg.tool_call_id);
-          parts2 = [{
-            functionResponse: {
-              name: ((_c2 = tool_call == null ? void 0 : tool_call.function) == null ? void 0 : _c2.name) || "unknown_tool",
-              response: { content: msg.content }
-            }
-          }];
-          if (options.responseFormat === "json" || ((_d2 = options.responseFormat) == null ? void 0 : _d2.type) === "json_schema" || options.responseSchema) {
-            parts2.push({ text: "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]" });
-          } else {
-            parts2.push({ text: "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]" });
-          }
-          break;
         default:
-          return null;
+          continue;
       }
-      return { role, parts: parts2 };
-    }).filter(Boolean);
+      contents.push({ role, parts: parts2 });
+    }
     while (contents.length > 0 && contents[0].role !== "user") {
       contents.shift();
     }
@@ -694,23 +762,17 @@ ${msg.content}`;
     if (tools && tools.length > 0) {
       requestOptions.config.tools = [{ functionDeclarations: tools.map((t) => t.function) }];
       if (requestOptions.config.responseMimeType === "application/json") {
-        console.warn("[GeminiProvider] Disabling strict JSON mode because tools are present. Relying on system prompt.");
+        console.warn(`[${this.constructor.name}] Disabling strict JSON mode because tools are present. Relying on system prompt.`);
         delete requestOptions.config.responseMimeType;
         delete requestOptions.config.responseSchema;
       }
     }
-    let response;
-    try {
-      response = await this.client.models.generateContent(requestOptions);
-    } catch (error) {
-      console.error(`[GeminiProvider] generateContent failed (API Key: ${this._getMaskedApiKey()}):`, error);
-      throw error;
-    }
-    const candidate = (_c = response.candidates) == null ? void 0 : _c[0];
+    const response = await this._generateContent(requestOptions);
+    const candidate = (_g = response.candidates) == null ? void 0 : _g[0];
     if (!candidate) {
       throw new LLMServiceException("No candidates returned from model", 500);
     }
-    const parts = ((_d = candidate.content) == null ? void 0 : _d.parts) || [];
+    const parts = ((_h = candidate.content) == null ? void 0 : _h.parts) || [];
     let textContent = "";
     let toolCalls = null;
     let responseThoughtSignature = null;
@@ -735,10 +797,8 @@ ${msg.content}`;
       }
     }
     if (!textContent && (!toolCalls || toolCalls.length === 0)) {
-      console.error("[GeminiProvider] Model returned empty response (no text, no tool calls)");
-      console.error("[GeminiProvider] Finish Reason:", candidate.finishReason);
-      console.error("[GeminiProvider] Safety Ratings:", JSON.stringify(candidate.safetyRatings, null, 2));
-      console.error("[GeminiProvider] Full Candidate:", JSON.stringify(candidate, null, 2));
+      console.error(`[${this.constructor.name}] Model returned empty response (no text, no tool calls)`);
+      console.error(`[${this.constructor.name}] Finish Reason:`, candidate.finishReason);
       throw new LLMServiceException(
         `Model returned empty response. Finish Reason: ${candidate.finishReason}.`,
         500
@@ -748,22 +808,18 @@ ${msg.content}`;
     return {
       content: textContent,
       thought_signature: responseThoughtSignature,
-      // Return signature to caller
       tool_calls: toolCalls ? (Array.isArray(toolCalls) ? toolCalls : [toolCalls]).map((fc) => ({
         type: "function",
         function: fc,
         thought_signature: fc.thought_signature
       })) : null,
       finishReason: normalizedFinishReason,
-      // Standardized: 'completed', 'truncated', etc.
       _rawFinishReason: candidate.finishReason,
-      // Keep original for debugging
       _responseFormat: options.responseFormat,
-      // Return usage stats
       usage: {
-        prompt_tokens: ((_e = response.usageMetadata) == null ? void 0 : _e.promptTokenCount) || 0,
-        completion_tokens: ((_f = response.usageMetadata) == null ? void 0 : _f.candidatesTokenCount) || 0,
-        total_tokens: ((_g = response.usageMetadata) == null ? void 0 : _g.totalTokenCount) || 0
+        prompt_tokens: ((_i = response.usageMetadata) == null ? void 0 : _i.promptTokenCount) || 0,
+        completion_tokens: ((_j = response.usageMetadata) == null ? void 0 : _j.candidatesTokenCount) || 0,
+        total_tokens: ((_k = response.usageMetadata) == null ? void 0 : _k.totalTokenCount) || 0
       },
       ...options.responseFormat && this._shouldAutoParse(options) ? {
         parsedContent: this._safeJsonParse(textContent)
@@ -785,7 +841,7 @@ ${msg.content}`;
         if (schema) {
           config.responseSchema = this._convertToGeminiSchema(schema);
         } else {
-          console.warn("[GeminiProvider] Using legacy JSON mode without schema - may produce markdown wrappers");
+          console.warn(`[${this.constructor.name}] Using legacy JSON mode without schema - may produce markdown wrappers`);
         }
       }
     }
@@ -843,8 +899,7 @@ ${msg.content}`;
     if (!content) return null;
     const parsed = extractJsonFromResponse(content);
     if (!parsed) {
-      console.error("[GeminiProvider] Failed to extract valid JSON from response");
-      console.error("[GeminiProvider] Content preview:", content.substring(0, 200));
+      console.error(`[${this.constructor.name}] Failed to extract valid JSON from response`);
     }
     return parsed;
   }
@@ -874,9 +929,9 @@ ${msg.content}`;
     toolResults.forEach((result) => messages.push({ role: "tool", tool_call_id: result.tool_call_id, content: result.output }));
   }
   async imageGeneration(prompt, systemPrompt, options = {}) {
-    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i;
     const modelName = options.model || this.models.image || "gemini-3-pro-image-preview";
-    console.log(`[GeminiProvider] Generating image with model: ${modelName}`);
+    console.log(`[${this.constructor.name}] Generating image with model: ${modelName}`);
     const hasReferenceImages = options.images && options.images.length > 0;
     const generationConfig = {
       responseModalities: hasReferenceImages ? ["TEXT", "IMAGE"] : ["IMAGE"]
@@ -908,7 +963,7 @@ ${msg.content}`;
     if (systemPrompt) {
       requestOptions.config.systemInstruction = { parts: [{ text: systemPrompt }] };
     }
-    const response = await this.client.models.generateContent(requestOptions);
+    const response = await this._generateContent(requestOptions);
     const imagePart = (_d = (_c = (_b = (_a = response.candidates) == null ? void 0 : _a[0]) == null ? void 0 : _b.content) == null ? void 0 : _c.parts) == null ? void 0 : _d.find(
       (part) => {
         var _a2;
@@ -916,30 +971,21 @@ ${msg.content}`;
       }
     );
     if (!imagePart || !imagePart.inlineData) {
-      const textPart = (_h = (_g = (_f = (_e = response.candidates) == null ? void 0 : _e[0]) == null ? void 0 : _f.content) == null ? void 0 : _g.parts) == null ? void 0 : _h.find((p) => p.text);
-      const candidate = (_i = response.candidates) == null ? void 0 : _i[0];
-      console.error("[GeminiProvider] Image generation failed (no image data)");
-      if (candidate) {
-        console.error("[GeminiProvider] Finish Reason:", candidate.finishReason);
-        console.error("[GeminiProvider] Safety Ratings:", JSON.stringify(candidate.safetyRatings, null, 2));
-        console.error("[GeminiProvider] Full Candidate:", JSON.stringify(candidate, null, 2));
-      }
-      if (textPart) {
-        console.warn("[GeminiProvider] Model returned text instead of image:", textPart.text);
-      }
+      const candidate = (_e = response.candidates) == null ? void 0 : _e[0];
+      console.error(`[${this.constructor.name}] Image generation failed (no image data)`);
       throw new Error(`No image data in response. Finish Reason: ${candidate == null ? void 0 : candidate.finishReason}`);
     }
     let thoughtSignature = null;
     if (imagePart.thought_signature || imagePart.thoughtSignature) {
       thoughtSignature = imagePart.thought_signature || imagePart.thoughtSignature;
     } else {
-      const signaturePart = (_m = (_l = (_k = (_j = response.candidates) == null ? void 0 : _j[0]) == null ? void 0 : _k.content) == null ? void 0 : _l.parts) == null ? void 0 : _m.find((p) => p.thought_signature || p.thoughtSignature);
+      const signaturePart = (_i = (_h = (_g = (_f = response.candidates) == null ? void 0 : _f[0]) == null ? void 0 : _g.content) == null ? void 0 : _h.parts) == null ? void 0 : _i.find((p) => p.thought_signature || p.thoughtSignature);
       if (signaturePart) {
         thoughtSignature = signaturePart.thought_signature || signaturePart.thoughtSignature;
       }
     }
     if (thoughtSignature && thoughtSignature.length > 5e4) {
-      console.warn(`[GeminiProvider] \u26A0\uFE0F  Thought signature is abnormally large (${thoughtSignature.length} chars). Replacing with bypass token to save context.`);
+      console.warn(`[${this.constructor.name}] \u26A0\uFE0F  Thought signature is abnormally large (${thoughtSignature.length} chars). Replacing with bypass token.`);
       thoughtSignature = "skip_thought_signature_validator";
     }
     return {
@@ -962,31 +1008,20 @@ ${prompt}` : prompt;
         durationSeconds: options.durationSeconds || 6,
         aspectRatio: options.aspectRatio || "16:9",
         numberOfVideos: 1,
-        // Pass reference images if provided
         ...images && images.length > 0 ? { referenceImages: images } : {}
       }
     };
-    const logConfig = JSON.parse(JSON.stringify(requestConfig));
-    if (logConfig.config && logConfig.config.referenceImages) {
-      logConfig.config.referenceImages = logConfig.config.referenceImages.map((img) => ({
-        ...img,
-        data: `... (${img.data ? img.data.length : 0} bytes)`
-        // Summarize data
-      }));
-    }
-    console.log("[GeminiProvider] startVideoGeneration request:", JSON.stringify(logConfig, null, 2));
     try {
       const operation = await this.client.models.generateVideos(requestConfig);
       this._pendingOperations.set(operation.name, operation);
       return { operationName: operation.name };
     } catch (error) {
-      console.error(`[GeminiProvider] startVideoGeneration failed (API Key: ${this._getMaskedApiKey()}):`, error);
+      console.error(`[${this.constructor.name}] startVideoGeneration failed (API Key: ${this._getMaskedApiKey()}):`, error);
       throw error;
     }
   }
   async getVideoGenerationStatus(operationName) {
     var _a, _b, _c, _d, _e, _f;
-    console.log(`[GeminiProvider] Checking status for operation: ${operationName}`);
     let operation = this._pendingOperations.get(operationName);
     if (!operation) {
       operation = await this.client.models.getOperation(operationName);
@@ -998,11 +1033,9 @@ ${prompt}` : prompt;
       progress: ((_a = operation.metadata) == null ? void 0 : _a.progressPercent) || 0,
       state: ((_b = operation.metadata) == null ? void 0 : _b.state) || (operation.done ? "COMPLETED" : "PROCESSING")
     };
-    console.log(`[GeminiProvider] Operation status: ${result.state}, Progress: ${result.progress}%`);
     if (operation.done) {
       this._pendingOperations.delete(operationName);
       if (operation.error) {
-        console.error("[GeminiProvider] Video generation failed:", JSON.stringify(operation.error, null, 2));
         result.error = operation.error;
       } else {
         const videoResult = operation.response;
@@ -1014,20 +1047,17 @@ ${prompt}` : prompt;
   }
   async startDeepResearch(prompt, options = {}) {
     const agent = options.agent || "deep-research-pro-preview-12-2025";
-    console.log(`[GeminiProvider] Starting Deep Research with agent: ${agent}`);
+    console.log(`[${this.constructor.name}] Starting Deep Research with agent: ${agent}`);
     try {
       const interaction = await this.client.interactions.create({
         agent,
         input: prompt,
         background: true,
-        // Required for long running
         store: true
-        // Required for polling
       });
-      console.log(`[GeminiProvider] Deep Research started. Interaction ID: ${interaction.id}`);
       return { operationId: interaction.id };
     } catch (error) {
-      console.error(`[GeminiProvider] startDeepResearch failed:`, error);
+      console.error(`[${this.constructor.name}] startDeepResearch failed:`, error);
       throw error;
     }
   }
@@ -1046,18 +1076,10 @@ ${prompt}` : prompt;
       }
       return response;
     } catch (error) {
-      console.error(`[GeminiProvider] getDeepResearchStatus failed for ${operationId}:`, error);
+      console.error(`[${this.constructor.name}] getDeepResearchStatus failed for ${operationId}:`, error);
       throw error;
     }
   }
-  /**
-   * Extract structured data from a file (PDF, Image, etc.) using Gemini Multimodal capabilities.
-   * @param {Buffer|string} fileData - Base64 string or Buffer of the file
-   * @param {string} mimeType - Mime type (e.g., 'application/pdf', 'image/png')
-   * @param {string} prompt - Extraction prompt
-   * @param {Object} schema - JSON schema for the output
-   * @param {Object} options - Additional options
-   */
   async extractWithLLM(fileData, mimeType, prompt, schema = null, options = {}) {
     var _a, _b, _c, _d;
     const tier = options.tier || "default";
@@ -1069,9 +1091,7 @@ ${prompt}` : prompt;
       maxTokens,
       temperature
     );
-    const parts = [
-      { text: prompt }
-    ];
+    const parts = [{ text: prompt }];
     let base64Data = fileData;
     if (typeof fileData !== "string") {
       try {
@@ -1094,7 +1114,7 @@ ${prompt}` : prompt;
       config: generationConfig
     };
     try {
-      const response = await this.client.models.generateContent(requestOptions);
+      const response = await this._generateContent(requestOptions);
       const candidate = (_a = response.candidates) == null ? void 0 : _a[0];
       if (!candidate) {
         throw new LLMServiceException("No candidates returned from model during extraction", 500);
@@ -1105,7 +1125,7 @@ ${prompt}` : prompt;
       }
       return textContent;
     } catch (error) {
-      console.error(`[GeminiProvider] extractWithLLM failed (API Key: ${this._getMaskedApiKey()}):`, error);
+      console.error(`[${this.constructor.name}] extractWithLLM failed (API Key: ${this._getMaskedApiKey()}):`, error);
       throw error;
     }
   }
@@ -1124,14 +1144,14 @@ var LLMService = class {
       return this.providerCache.get(cacheKey);
     }
     const config = await ConfigManager.getConfig(tenantId, this.env);
-    if (!config.apiKey) {
+    if (!config.apiKey && config.provider !== "vertex") {
       throw new LLMServiceException(`LLM service is not configured for ${config.provider}. Missing API Key.`, 500);
     }
     let provider;
     if (config.provider === "openai") {
       provider = new OpenAIProvider(config);
-    } else if (config.provider === "gemini") {
-      provider = new GeminiProvider(config);
+    } else if (config.provider === "gemini" || config.provider === "vertex") {
+      provider = new GoogleProvider(config);
     } else {
       throw new LLMServiceException(`Unsupported LLM provider: ${config.provider}`, 500);
     }
@@ -1703,13 +1723,15 @@ export {
   ConfigManager,
   DefaultConfigProvider,
   FINISH_REASONS,
-  GeminiProvider,
+  GoogleProvider as GeminiProvider,
+  GoogleProvider,
   LLMService,
   LLMServiceException,
   MODEL_CONFIGS,
   OpenAIProvider,
   TranscriptionService,
   TranscriptionServiceException,
+  GoogleProvider as VertexProvider,
   createSpeechHandler,
   extractJsonFromResponse,
   extractTextAndJson,