@contentgrowth/llm-service 1.0.9 → 1.1.1

package/dist/index.cjs

@@ -442,7 +442,7 @@ var OpenAIProvider = class extends BaseLLMProvider {
     );
   }
   async _chatCompletionWithModel(messages, systemPrompt, tools, modelName, maxTokens, temperature, options = {}) {
-    var _a, _b;
+    var _a, _b, _c, _d, _e;
     const requestPayload = {
       model: modelName,
       temperature: (_a = options.temperature) != null ? _a : temperature,
@@ -479,11 +479,19 @@ var OpenAIProvider = class extends BaseLLMProvider {
       _rawFinishReason: rawFinishReason,
       // Keep original for debugging
       // Add metadata about response format
+      // Add metadata about response format
       _responseFormat: options.responseFormat,
+      // Return usage stats
+      usage: {
+        prompt_tokens: ((_c = response.usage) == null ? void 0 : _c.prompt_tokens) || 0,
+        completion_tokens: ((_d = response.usage) == null ? void 0 : _d.completion_tokens) || 0,
+        total_tokens: ((_e = response.usage) == null ? void 0 : _e.total_tokens) || 0
+      },
       // Auto-parse JSON if requested
       ...options.responseFormat && this._shouldAutoParse(options) ? {
         parsedContent: this._safeJsonParse(message.content)
-      } : {}
+      } : {},
+      model: modelName
     };
   }
   _buildResponseFormat(options) {
@@ -626,7 +634,7 @@ var GeminiProvider = class extends BaseLLMProvider {
     );
   }
   async _chatCompletionWithModel(messages, systemPrompt, tools, modelName, maxTokens, temperature, options = {}) {
-    var _a, _b, _c, _d;
+    var _a, _b, _c, _d, _e, _f, _g;
     const generationConfig = {
       temperature: (_a = options.temperature) != null ? _a : temperature,
       maxOutputTokens: (_b = options.maxTokens) != null ? _b : maxTokens
@@ -684,7 +692,6 @@ ${msg.content}`;
                 functionCall: { name: tc.function.name, args: tc.function.arguments || tc.function.args }
               };
               if (tc.thought_signature) {
-                console.log(`[GeminiProvider] Sending thought_signature in tool_call (${tc.thought_signature.length} chars)`);
                 part.thoughtSignature = tc.thought_signature;
               }
               return part;
@@ -692,7 +699,6 @@ ${msg.content}`;
           } else {
             const part = { text: msg.content || "" };
             if (isLastAssistantMessage && msg.thought_signature) {
-              console.log(`[GeminiProvider] Sending thought_signature in text message (${msg.thought_signature.length} chars)`);
               part.thoughtSignature = msg.thought_signature;
             }
             parts2 = [part];
@@ -801,9 +807,16 @@ ${msg.content}`;
       _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
+      },
       ...options.responseFormat && this._shouldAutoParse(options) ? {
         parsedContent: this._safeJsonParse(textContent)
-      } : {}
+      } : {},
+      model: modelName
     };
   }
   _buildGenerationConfig(options, maxTokens, temperature) {
@@ -890,12 +903,15 @@ ${msg.content}`;
         const tool = toolImplementations[toolName];
         const tool_call_id = `gemini-tool-call-${index}`;
         toolCall.id = tool_call_id;
+        console.log(`[Tool Call] > ${toolName}:`, JSON.stringify(toolCall.function.args));
         if (!tool) {
           console.error(`[Tool Error] Tool '${toolName}' not found`);
           return { tool_call_id, output: JSON.stringify({ error: `Tool '${toolName}' not found.` }) };
         }
         try {
           const output = await tool(toolCall.function.args, { env, tenantId });
+          const preview = typeof output === "string" ? output.length > 200 ? output.substring(0, 200) + "..." : output : JSON.stringify(output).substring(0, 200) + "...";
+          console.log(`[Tool Result] < ${toolName}: ${preview}`);
           return { tool_call_id, output };
         } catch (error) {
           console.error(`[Tool Error] ${toolName} failed:`, error.message);
@@ -909,8 +925,9 @@ ${msg.content}`;
     var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m;
     const modelName = options.model || this.models.image || "gemini-3-pro-image-preview";
     console.log(`[GeminiProvider] Generating image with model: ${modelName}`);
+    const hasReferenceImages = options.images && options.images.length > 0;
     const generationConfig = {
-      responseModalities: ["IMAGE"]
+      responseModalities: hasReferenceImages ? ["TEXT", "IMAGE"] : ["IMAGE"]
     };
     if (options.aspectRatio) {
       generationConfig.imageConfig = {
@@ -918,7 +935,7 @@ ${msg.content}`;
       };
     }
     const parts = [{ text: prompt }];
-    if (options.images && options.images.length > 0) {
+    if (hasReferenceImages) {
       options.images.forEach((img) => {
         parts.push({
           inlineData: {
@@ -1264,16 +1281,26 @@ var LLMService = class {
     let currentMessages = [...messages];
     const MAX_ITERATIONS = 10;
     let iteration = 0;
+    const accumulateUsage = (responseUsage) => {
+      if (responseUsage) {
+        totalUsage.prompt_tokens += responseUsage.prompt_tokens || 0;
+        totalUsage.completion_tokens += responseUsage.completion_tokens || 0;
+        totalUsage.total_tokens += responseUsage.total_tokens || 0;
+      }
+    };
     const initialResponse = await provider.chatCompletion(
       currentMessages,
       systemPrompt,
       tools,
       options
     );
+    const totalUsage = { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 };
+    accumulateUsage(initialResponse.usage);
     let { content, tool_calls, parsedContent, finishReason, _rawFinishReason } = initialResponse;
     while (tool_calls && iteration < MAX_ITERATIONS) {
       iteration++;
-      console.log(`[Tool Call] Iteration ${iteration}/${MAX_ITERATIONS} with finish reason ${finishReason}: Assistant wants to use tools:`, tool_calls);
+      const toolNames = tool_calls.map((tc) => tc.function.name).join(", ");
+      console.log(`[Tool Call] Iteration ${iteration}/${MAX_ITERATIONS}: Using tools - [${toolNames}]`);
       currentMessages.push({ role: "assistant", content: content || "", tool_calls });
       await provider.executeTools(tool_calls, currentMessages, tenantId, this.toolImplementations, this.env);
       const nextResponse = await provider.chatCompletion(
@@ -1282,6 +1309,7 @@ var LLMService = class {
         tools,
         options
       );
+      accumulateUsage(nextResponse.usage);
       content = nextResponse.content;
       tool_calls = nextResponse.tool_calls;
       parsedContent = nextResponse.parsedContent;
@@ -1291,7 +1319,7 @@ var LLMService = class {
     if (iteration >= MAX_ITERATIONS) {
       console.warn(`[Tool Call] Reached maximum iterations (${MAX_ITERATIONS}). Forcing completion.`);
     }
-    return { content, parsedContent, toolCalls: tool_calls, finishReason, _rawFinishReason };
+    return { content, parsedContent, toolCalls: tool_calls, finishReason, _rawFinishReason, usage: totalUsage, model: initialResponse.model || "unknown-model" };
   }
   /**
    * Generate a video (async wrapper with polling - backward compatibility)