fix(ai): isolate heavy Ollama workloads from GCP alert lane

apps/api/src/services/knowledge_rag_service.py

@@ -20,11 +20,10 @@ from pathlib import Path
 import httpx
 import structlog
 
-from src.core.config import get_settings
 import src.repositories.rag_chunk_repository as rag_repo
+from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
-settings = get_settings()
 
 _EMBED_MODEL = "nomic-embed-text"
 _GEN_MODEL = "qwen2.5:7b-instruct"
@@ -131,7 +130,7 @@ class KnowledgeRAGService:
         try:
             http = await self._get_http()
             resp = await http.post(
-                f"{settings.OLLAMA_URL}/api/embeddings",
+                f"{resolve_ollama_endpoint('embedding')}/api/embeddings",
                 json={"model": _EMBED_MODEL, "prompt": text},
             )
             if resp.status_code == 200:
@@ -150,7 +149,7 @@ class KnowledgeRAGService:
         try:
             http = await self._get_http()
             resp = await http.post(
-                f"{settings.OLLAMA_URL}/api/generate",
+                f"{resolve_ollama_endpoint('rag')}/api/generate",
                 json={
                     "model": _GEN_MODEL,
                     "prompt": prompt,