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

apps/api/src/services/knowledge_rag_service.py

@@ -1,13 +1,13 @@
 """
 AWOOOI — Knowledge RAG Service (Phase 33, ADR-067)
 ==================================================
-本地 RAG 知識庫：nomic-embed-text 768維向量 + pgvector
+本地 RAG 知識庫：bge-m3 1024維向量 + pgvector
 
 索引策略:
   - 初期 < 100 筆: 線性搜尋
   - 超過 100 筆: 執行 CREATE INDEX ivfflat (手動觸發)
 
-向量模型: nomic-embed-text (Ollama 111, 768維) — 188:11434 被 NetworkPolicy v1.3 封閉
+向量模型: bge-m3 (GCP-A/GCP-B/111 Ollama lane, 1024維)
 生成模型: qwen2.5:7b-instruct (Ollama 111)
 
 leWOOOgo: Service 層只處理業務邏輯，DB 存取委派 rag_chunk_repository
@@ -21,11 +21,12 @@ import httpx
 import structlog
 
 import src.repositories.rag_chunk_repository as rag_repo
+from src.core.config import settings
 from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
 
-_EMBED_MODEL = "nomic-embed-text"
+_EMBED_MODEL = "bge-m3:latest"
 _GEN_MODEL = "qwen2.5:7b-instruct"
 _TOP_K = 5
 
@@ -131,7 +132,10 @@ class KnowledgeRAGService:
             http = await self._get_http()
             resp = await http.post(
                 f"{resolve_ollama_endpoint('embedding')}/api/embeddings",
-                json={"model": _EMBED_MODEL, "prompt": text},
+                json={
+                    "model": getattr(settings, "OLLAMA_EMBEDDING_MODEL", _EMBED_MODEL),
+                    "prompt": text,
+                },
             )
             if resp.status_code == 200:
                 return resp.json().get("embedding")