@contentgrowth/llm-service 0.7.0 → 0.7.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@contentgrowth/llm-service",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "description": "Unified LLM Service for Content Growth",
   "main": "src/index.js",
   "type": "module",

package/src/llm/providers/gemini-provider.js

@@ -90,6 +90,26 @@ export class GeminiProvider extends BaseLLMProvider {
                 case 'user':
                     role = 'user';
                     parts = [{ text: msg.content }];
+
+                    // Enhancement: If this is the LAST message (current turn), append the reminder.
+                    // This helps the model respect the system prompt (especially format) even with long context history.
+                    if (index === geminiMessages.length - 1) {
+                        let reminder = "";
+                        if (options.responseFormat === 'json' || options.responseFormat?.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.]";
+                        }
+
+                        // Append to the existing text part (Safest method)
+                        const lastPart = parts.find(p => p.text);
+                        if (lastPart) {
+                            lastPart.text += reminder;
+                        } else {
+                            // Fallback if message was image-only
+                            parts.push({ text: reminder });
+                        }
+                    }
                     break;
                 case 'assistant':
                     role = 'model';
@@ -115,11 +135,11 @@ export class GeminiProvider extends BaseLLMProvider {
                     // Fix for JSON mode: If JSON is requested, remind the model to output JSON after tool execution
                     // This is necessary because strict JSON mode is disabled when tools are present.
                     if (options.responseFormat === 'json' || options.responseFormat?.type === 'json_schema' || options.responseSchema) {
-                        parts.push({ text: "Please ensure the final response is valid JSON as per the system instructions." });
+                        parts.push({ text: "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]" });
                     } else {
                         // Generic reminder to help model stay on track with system prompt instructions (e.g. formatting)
                         // even if no specific JSON mode is configured.
-                        parts.push({ text: "Please ensure the final response follows the system prompt instructions." });
+                        parts.push({ text: "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]" });
                     }
                     break;
                 default:
@@ -160,7 +180,7 @@ export class GeminiProvider extends BaseLLMProvider {
             }
         }
 
-        console.log('[GeminiProvider] generateContent request:', JSON.stringify(requestOptions, null, 2));
+        // console.log('[GeminiProvider] generateContent request:', JSON.stringify(requestOptions, null, 2));
 
         let response;
         try {
@@ -208,7 +228,7 @@ export class GeminiProvider extends BaseLLMProvider {
             );
         }
 
-        console.log('Gemini returns:', textContent);
+        // console.log('Gemini returns:', textContent);
         // Return with parsed JSON if applicable
         return {
             content: textContent,
@@ -245,7 +265,7 @@ export class GeminiProvider extends BaseLLMProvider {
                     // Use responseSchema for strict structured output
                     // Must convert to Gemini Schema format (Uppercase types)
                     config.responseSchema = this._convertToGeminiSchema(schema);
-                    console.log('[GeminiProvider] Using Strict JSON mode with schema (responseSchema)');
+                    // console.log('[GeminiProvider] Using Strict JSON mode with schema (responseSchema)');
                 } else {
                     console.warn('[GeminiProvider] Using legacy JSON mode without schema - may produce markdown wrappers');
                 }
@@ -316,9 +336,7 @@ export class GeminiProvider extends BaseLLMProvider {
         // - Brace extraction as fallback
         const parsed = extractJsonFromResponse(content);
 
-        if (parsed) {
-            console.log('[GeminiProvider] Successfully parsed JSON from response');
-        } else {
+        if (!parsed) {
             console.error('[GeminiProvider] Failed to extract valid JSON from response');
             console.error('[GeminiProvider] Content preview:', content.substring(0, 200));
         }
@@ -334,7 +352,7 @@ export class GeminiProvider extends BaseLLMProvider {
                 const tool_call_id = `gemini-tool-call-${index}`;
                 toolCall.id = tool_call_id;
 
-                console.log(`[Tool Call] ${toolName} with arguments:`, toolCall.function.args);
+                // console.log(`[Tool Call] ${toolName} with arguments:`, toolCall.function.args);
 
                 if (!tool) {
                     console.error(`[Tool Error] Tool '${toolName}' not found`);
@@ -342,7 +360,7 @@ export class GeminiProvider extends BaseLLMProvider {
                 }
                 try {
                     const output = await tool(toolCall.function.args, { env, tenantId });
-                    console.log(`[Tool Result] ${toolName} returned:`, output.substring(0, 200) + (output.length > 200 ? '...' : ''));
+                    // console.log(`[Tool Result] ${toolName} returned:`, output.substring(0, 200) + (output.length > 200 ? '...' : ''));
                     return { tool_call_id, output };
                 } catch (error) {
                     console.error(`[Tool Error] ${toolName} failed:`, error.message);
@@ -391,7 +409,7 @@ export class GeminiProvider extends BaseLLMProvider {
             requestOptions.config.systemInstruction = { parts: [{ text: systemPrompt }] };
         }
 
-        console.log('[GeminiProvider] imageGeneration request:', JSON.stringify(requestOptions, null, 2));
+        // console.log('[GeminiProvider] imageGeneration request:', JSON.stringify(requestOptions, null, 2));
 
         const response = await this.client.models.generateContent(requestOptions);
 
@@ -429,9 +447,12 @@ export class GeminiProvider extends BaseLLMProvider {
 
     async startVideoGeneration(prompt, images, modelName, systemPrompt, options = {}) {
         // Use unified client for video generation
-        const operation = await this.client.models.generateVideos({
+        // Prepend system prompt to user prompt if provided, as video models often expect instructions in the prompt
+        const effectivePrompt = systemPrompt ? `${systemPrompt}\n\n${prompt}` : prompt;
+
+        const requestConfig = {
             model: modelName,
-            prompt: prompt,
+            prompt: effectivePrompt,
             config: {
                 durationSeconds: options.durationSeconds || 6,
                 aspectRatio: options.aspectRatio || '16:9',
@@ -439,15 +460,35 @@ export class GeminiProvider extends BaseLLMProvider {
                 // Pass reference images if provided
                 ...(images && images.length > 0 ? { referenceImages: images } : {}),
             }
-        });
+        };
+
+        // Create a loggable copy of the config
+        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));
 
-        // Store operation for later polling
-        this._pendingOperations.set(operation.name, operation);
+        try {
+            const operation = await this.client.models.generateVideos(requestConfig);
 
-        return { operationName: operation.name };
+            // Store operation for later polling
+            this._pendingOperations.set(operation.name, operation);
+
+            return { operationName: operation.name };
+        } catch (error) {
+            console.error('[GeminiProvider] startVideoGeneration failed:', error);
+            throw error;
+        }
     }
 
     async getVideoGenerationStatus(operationName) {
+        console.log(`[GeminiProvider] Checking status for operation: ${operationName}`);
+
         // Get the operation from cache or fetch it
         let operation = this._pendingOperations.get(operationName);
 
@@ -468,6 +509,8 @@ export class GeminiProvider extends BaseLLMProvider {
             state: operation.metadata?.state || (operation.done ? 'COMPLETED' : 'PROCESSING'),
         };
 
+        console.log(`[GeminiProvider] Operation status: ${result.state}, Progress: ${result.progress}%`);
+
         if (operation.done) {
             // Clean up from cache
             this._pendingOperations.delete(operationName);

package/src/llm-service.js

@@ -159,18 +159,19 @@ export class LLMService {
         }
     }
 
-    async chatWithTools(messages, tenantId, systemPrompt, tools = []) {
+    async chatWithTools(messages, tenantId, systemPrompt, tools = [], options = {}) {
         const provider = await this._getProvider(tenantId);
 
         let currentMessages = [...messages];
         const MAX_ITERATIONS = 10; // Prevent infinite loops
         let iteration = 0;
 
-        // Initial call - no options since tools + JSON mode are incompatible
+        // Initial call
         const initialResponse = await provider.chatCompletion(
             currentMessages,
             systemPrompt,
-            tools
+            tools,
+            options
         );
 
         let { content, tool_calls, parsedContent } = initialResponse;
@@ -184,11 +185,12 @@ export class LLMService {
             // Execute tools using the provider's helper (which formats results for that provider)
             await provider.executeTools(tool_calls, currentMessages, tenantId, this.toolImplementations, this.env);
 
-            // Next call - no options
+            // Next call
             const nextResponse = await provider.chatCompletion(
                 currentMessages,
                 systemPrompt,
-                tools
+                tools,
+                options
             );
 
             content = nextResponse.content;