feat: Add curated expert occupations with local data sources

- Add curated occupations seed files (210 entries in zh/en) with specific domains
- Add DBpedia occupations data (2164 entries) for external source option
- Refactor expert_source_service to read from local JSON files
- Improve keyword generation prompts to leverage expert domain context
- Add architecture analysis documentation (ARCHITECTURE_ANALYSIS.md)
- Fix expert source selection bug (proper handling of empty custom_experts)
- Update frontend to support curated/dbpedia/wikidata expert sources

Key changes:
- backend/app/data/: Local occupation data files
- backend/app/services/expert_source_service.py: Simplified local file reading
- backend/app/prompts/expert_transformation_prompt.py: Better domain-aware prompts
- Removed expert_cache.py (no longer needed with local files)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

backend/app/routers/expert_transformation.py

@@ -37,16 +37,29 @@ async def generate_expert_transformation_events(
         model = request.model
 
         # ========== Step 0: Generate expert team ==========
+        logger.info(f"[DEBUG] expert_source from request: {request.expert_source}")
+        logger.info(f"[DEBUG] expert_source value: {request.expert_source.value}")
+        logger.info(f"[DEBUG] custom_experts: {request.custom_experts}")
+
         yield f"event: expert_start\ndata: {json.dumps({'message': '正在組建專家團隊...', 'source': request.expert_source.value}, ensure_ascii=False)}\n\n"
 
         experts: List[ExpertProfile] = []
         actual_source = request.expert_source.value
 
+        # 過濾出實際有內容的自訂專家（排除空字串）
+        actual_custom_experts = [
+            e.strip() for e in (request.custom_experts or [])
+            if e and e.strip()
+        ]
+        logger.info(f"[DEBUG] actual_custom_experts (filtered): {actual_custom_experts}")
+
         # 決定使用哪種來源生成專家
+        # 只有在明確選擇 LLM 或有實際自訂專家時才使用 LLM
         use_llm = (
             request.expert_source == ExpertSource.LLM or
-            request.custom_experts  # 有自訂專家時，使用 LLM 補充
+            len(actual_custom_experts) > 0  # 有實際自訂專家時，使用 LLM 補充
         )
+        logger.info(f"[DEBUG] use_llm decision: {use_llm}")
 
         if use_llm:
             # LLM 生成專家
@@ -55,7 +68,7 @@ async def generate_expert_transformation_events(
                     query=request.query,
                     categories=all_categories,
                     expert_count=request.expert_count,
-                    custom_experts=request.custom_experts
+                    custom_experts=actual_custom_experts if actual_custom_experts else None
                 )
                 logger.info(f"Expert prompt: {expert_prompt[:200]}")
 
@@ -78,9 +91,9 @@ async def generate_expert_transformation_events(
                 yield f"event: error\ndata: {json.dumps({'error': f'專家團隊生成失敗: {str(e)}'}, ensure_ascii=False)}\n\n"
                 return
         else:
-            # 外部來源生成專家
+            # 外部來源生成專家 (本地檔案，同步)
             try:
-                experts_data, actual_source = await expert_source_service.get_experts(
+                experts_data, actual_source = expert_source_service.get_experts(
                     source=request.expert_source.value,
                     count=request.expert_count,
                     language=request.expert_language
@@ -106,7 +119,7 @@ async def generate_expert_transformation_events(
                         query=request.query,
                         categories=all_categories,
                         expert_count=request.expert_count,
-                        custom_experts=request.custom_experts
+                        custom_experts=actual_custom_experts if actual_custom_experts else None
                     )
 
                     expert_response = await ollama_provider.generate(