fix: AI LLM backup model not producing content; add retry for thinking models

- Replace backup model stepfun-ai/step-3.5-flash with meta/llama-3.1-8b-instruct
  (stepfun is a thinking model that uses all tokens on reasoning and
  never outputs content, causing all 3 fallthroughs to fail)
- Add retry with doubled max_tokens when primary model returns empty
  content (deepseek-v4-flash thinking can exhaust token budget)
- Increase backup timeout to 120s and max_tokens to min 2048
- Move callApi error handling to return null instead of throw for
  cleaner fallthrough logic with timeout logging

backend/src/modules/ai/ai.service.ts

@@ -22,44 +22,47 @@ export class AiService {
 
   private readonly backupUrl = process.env.AI_BACKUP_URL || "https://integrate.api.nvidia.com/v1"
   private readonly backupKey = process.env.AI_BACKUP_KEY || ""
-  private readonly backupModel = process.env.AI_BACKUP_MODEL || "stepfun-ai/step-3.5-flash"
+  private readonly backupModel = process.env.AI_BACKUP_MODEL || "meta/llama-3.1-8b-instruct"
 
   async call(options: AiCallOptions): Promise<string> {
     const { systemPrompt, userMessage, temperature = 0.7, maxTokens = 2048 } = options
 
     // Try primary AI (deepseek-v4-flash on sensenova)
     try {
-      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, maxTokens, 60000)
       if (result) return result
+      // Primary returned empty content (thinking model exhausted tokens); retry with more tokens
+      const retry = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, Math.min(maxTokens * 2, 4096), 60000)
+      if (retry) return retry
     } catch (e) {
       this.logger.warn(`Primary AI failed: ${(e as Error).message}, trying primary fallback...`)
     }
 
     // Try primary fallback model (sensenova-6.7-flash-lite, same provider)
     try {
-      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryFallbackModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryFallbackModel, systemPrompt, userMessage, temperature, maxTokens, 60000)
       if (result) return result
     } catch (e) {
       this.logger.warn(`Primary fallback AI also failed: ${(e as Error).message}, trying backup...`)
     }
 
-    // Try backup AI (NVIDIA)
+    // Try backup AI (NVIDIA - meta/llama-3.1-8b-instruct)
     try {
-      const result = await this.callApi(this.backupUrl, this.backupKey, this.backupModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.backupUrl, this.backupKey, this.backupModel, systemPrompt, userMessage, temperature, Math.max(maxTokens, 2048), 120000)
       if (result) return result
     } catch (e) {
       this.logger.warn(`Backup AI also failed: ${(e as Error).message}`)
     }
 
-    // Final fallback
+    throw new Error("AI 服务暂时不可用，请稍后重试")
-    throw new Error("AI \u670d\u52a1\u6682\u65f6\u4e0d\u53ef\u7528\uff0c\u8bf7\u7a0d\u540e\u91cd\u8bd5")
   }
 
   private async callApi(
     baseUrl: string, apiKey: string, model: string,
     systemPrompt: string, userMessage: string,
-    temperature: number, maxTokens: number,
+    temperature: number, maxTokens: number, timeout: number,
   ): Promise<string | null> {
+    try {
       const res = await axios.post(
         `${baseUrl}/chat/completions`,
         {
@@ -76,11 +79,17 @@ export class AiService {
             "Authorization": `Bearer ${apiKey}`,
             "Content-Type": "application/json",
           },
-        timeout: 60000,
+          timeout,
           httpsAgent: httpAgent,
           transitional: { clarifyTimeoutError: true },
         },
       )
       return res.data?.choices?.[0]?.message?.content || null
+    } catch (e: any) {
+      if (e.code === 'ECONNABORTED') {
+        this.logger.warn(`AI call timeout (${timeout}ms): ${model}`)
+      }
+      return null
+    }
   }
 }