@contentgrowth/llm-service 0.7.2 → 0.7.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@contentgrowth/llm-service",
-  "version": "0.7.2",
+  "version": "0.7.3",
   "description": "Unified LLM Service for Content Growth",
   "main": "src/index.js",
   "type": "module",

package/src/index.js

@@ -5,3 +5,5 @@ export { MODEL_CONFIGS } from './llm/config-manager.js';
 export { OpenAIProvider } from './llm/providers/openai-provider.js';
 export { GeminiProvider } from './llm/providers/gemini-provider.js';
 export { extractJsonFromResponse } from './llm/json-utils.js';
+export { FINISH_REASONS } from './llm/providers/base-provider.js';
+

package/src/llm/config-provider.js

@@ -94,7 +94,7 @@ export class DefaultConfigProvider extends BaseConfigProvider {
             apiKey: tenantConfig.api_key,
             models: MODEL_CONFIGS[tenantConfig.provider],
             temperature: parseFloat(env.DEFAULT_TEMPERATURE || '0.7'),
-            maxTokens: parseInt(env.DEFAULT_MAX_TOKENS || '4096'),
+            maxTokens: parseInt(env.DEFAULT_MAX_TOKENS || '16384'),
             capabilities: tenantConfig.capabilities || { chat: true, image: false, video: false },
             isTenantOwned: true
         };

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

@@ -1,3 +1,15 @@
+/**
+ * Standardized finish reasons across all LLM providers.
+ * Providers map their native values to these standard constants.
+ */
+export const FINISH_REASONS = {
+    COMPLETED: 'completed',      // Normal completion (OpenAI: stop, Gemini: STOP, Anthropic: end_turn)
+    TRUNCATED: 'truncated',      // Hit max tokens (OpenAI: length, Gemini: MAX_TOKENS, Anthropic: max_tokens)
+    CONTENT_FILTER: 'content_filter', // Content was filtered
+    TOOL_CALL: 'tool_call',      // Stopped for tool call
+    UNKNOWN: 'unknown',          // Unknown/unmapped reason
+};
+
 /**
  * Abstract base class for LLM Providers.
  * Defines the standard interface that all providers must implement.
@@ -7,6 +19,39 @@ export class BaseLLMProvider {
         this.config = config;
     }
 
+    /**
+     * Normalize provider-specific finish reason to standard value.
+     * Override in subclass if provider uses different values.
+     * @param {string} providerReason - The provider's native finish reason
+     * @returns {string} Standardized finish reason from FINISH_REASONS
+     */
+    normalizeFinishReason(providerReason) {
+        // Default mappings - providers can override
+        const upperReason = (providerReason || '').toUpperCase();
+
+        // Completed mappings
+        if (['STOP', 'END_TURN'].includes(upperReason)) {
+            return FINISH_REASONS.COMPLETED;
+        }
+
+        // Truncated mappings
+        if (['LENGTH', 'MAX_TOKENS'].includes(upperReason)) {
+            return FINISH_REASONS.TRUNCATED;
+        }
+
+        // Content filter mappings
+        if (['CONTENT_FILTER', 'SAFETY'].includes(upperReason)) {
+            return FINISH_REASONS.CONTENT_FILTER;
+        }
+
+        // Tool call mappings
+        if (['TOOL_CALLS', 'TOOL_USE', 'FUNCTION_CALL'].includes(upperReason)) {
+            return FINISH_REASONS.TOOL_CALL;
+        }
+
+        return FINISH_REASONS.UNKNOWN;
+    }
+
     /**
      * Simple chat interface for single-turn conversations
      * @param {string} userMessage 

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

@@ -230,9 +230,14 @@ export class GeminiProvider extends BaseLLMProvider {
 
         // console.log('Gemini returns:', textContent);
         // Return with parsed JSON if applicable
+        // Normalize the finish reason to standard value for consistent handling
+        const normalizedFinishReason = this.normalizeFinishReason(candidate.finishReason);
+
         return {
             content: textContent,
             tool_calls: toolCalls ? (Array.isArray(toolCalls) ? toolCalls : [toolCalls]).map(fc => ({ type: 'function', function: fc })) : null,
+            finishReason: normalizedFinishReason, // Standardized: 'completed', 'truncated', etc.
+            _rawFinishReason: candidate.finishReason, // Keep original for debugging
             _responseFormat: options.responseFormat,
             ...(options.responseFormat && this._shouldAutoParse(options) ? {
                 parsedContent: this._safeJsonParse(textContent)

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

@@ -68,10 +68,16 @@ export class OpenAIProvider extends BaseLLMProvider {
             );
         }
 
+        // Normalize the finish reason to standard value for consistent handling
+        const rawFinishReason = response.choices[0].finish_reason;
+        const normalizedFinishReason = this.normalizeFinishReason(rawFinishReason);
+
         // Return with parsed JSON if applicable
         return {
             content: message.content,
             tool_calls: message.tool_calls,
+            finishReason: normalizedFinishReason, // Standardized: 'completed', 'truncated', etc.
+            _rawFinishReason: rawFinishReason, // Keep original for debugging
             // Add metadata about response format
             _responseFormat: options.responseFormat,
             // Auto-parse JSON if requested

package/src/llm-service.js

@@ -174,12 +174,12 @@ export class LLMService {
             options
         );
 
-        let { content, tool_calls, parsedContent } = initialResponse;
+        let { content, tool_calls, parsedContent, finishReason } = initialResponse;
 
         // Tool execution loop with safety limit
         while (tool_calls && iteration < MAX_ITERATIONS) {
             iteration++;
-            console.log(`[Tool Call] Iteration ${iteration}/${MAX_ITERATIONS}: Assistant wants to use tools:`, tool_calls);
+            console.log(`[Tool Call] Iteration ${iteration}/${MAX_ITERATIONS} with finish reason ${finishReason}: Assistant wants to use tools:`, tool_calls);
             currentMessages.push({ role: 'assistant', content: content || '', tool_calls });
 
             // Execute tools using the provider's helper (which formats results for that provider)
@@ -196,6 +196,7 @@ export class LLMService {
             content = nextResponse.content;
             tool_calls = nextResponse.tool_calls;
             parsedContent = nextResponse.parsedContent; // Preserve parsedContent from final response
+            finishReason = nextResponse.finishReason; // Preserve finishReason from final response
         }
 
         if (iteration >= MAX_ITERATIONS) {
@@ -203,7 +204,7 @@ export class LLMService {
         }
 
         // Return both content and parsedContent (if available)
-        return { content, parsedContent, toolCalls: tool_calls };
+        return { content, parsedContent, toolCalls: tool_calls, finishReason };
     }
 
     /**