fix(rag): use bge embeddings on GCP Ollama lane

apps/api/src/services/playbook_rag.py

@@ -4,7 +4,7 @@ Playbook RAG Service - Phase 3 向量化語意搜尋
 ADR-030: 智能自動修復系統
 
 使用 Embedding 進行 Playbook 語意搜尋:
-1. Ollama nomic-embed-text 生成向量
+1. Ollama bge-m3 生成向量
 2. Redis 儲存向量 (JSON 格式)
 3. 餘弦相似度搜尋
 
@@ -41,9 +41,9 @@ logger = structlog.get_logger(__name__)
 # Constants
 # =============================================================================
 
-# Embedding Model (Ollama 本地)
-EMBEDDING_MODEL = "nomic-embed-text"
-EMBEDDING_DIM = 768  # nomic-embed-text 向量維度
+# Embedding Model (Ollama)
+EMBEDDING_MODEL = "bge-m3:latest"
+EMBEDDING_DIM = 1024  # bge-m3 向量維度
 
 
 def _dedupe_urls(urls: list[str]) -> list[str]:
@@ -170,7 +170,7 @@ class PlaybookRAGService:
                 getattr(settings, "OLLAMA_FALLBACK_URL", ""),
             ]
         )
-        self.embedding_model = EMBEDDING_MODEL
+        self.embedding_model = str(getattr(settings, "OLLAMA_EMBEDDING_MODEL", EMBEDDING_MODEL) or EMBEDDING_MODEL)
 
     # =========================================================================
     # Embedding Operations