Update settings.py

settings.py

@@ -1,29 +1,36 @@
 import os
 from typing import Dict, Any, List
 
-# Residency / deployment guidance (enforce in your infra)
+# ===== Residency / Deployment =====
 CANADA_RESIDENCY_REQUIRED = True
-CLOUD_REGION = os.getenv("CLOUD_REGION", "ca-central-1")
+CLOUD_REGION = os.getenv("CLOUD_REGION", "ca")  # e.g., "ca" or "us"
 
-# Keys
+# ===== Cohere API =====
 COHERE_API_KEY = os.getenv("COHERE_API_KEY", "")
+# Optional: override Cohere API url/region if your enterprise deployment provides one
+COHERE_API_URL = os.getenv("COHERE_API_URL", "")  # leave blank to use default SDK endpoint
+COHERE_TIMEOUT_S = float(os.getenv("COHERE_TIMEOUT_S", "45"))
 
-# Cohere models (ask Cohere for larger/custom healthcare)
+# Cohere models (ask Cohere for larger/custom healthcare models as needed)
 COHERE_MODEL_PRIMARY = os.getenv("COHERE_MODEL_PRIMARY", "command-r")
 COHERE_EMBED_MODEL = os.getenv("COHERE_EMBED_MODEL", "embed-english-v3.0")
 
-# Open-model fallbacks (largest first; will try in order)
+# ===== Fallbacks (OPEN MODELS ONLY) =====
+# By request, we default to Cohere-only. Set to "1" to allow open-model fallbacks.
+USE_OPEN_FALLBACKS = os.getenv("USE_OPEN_FALLBACKS", "0") in ("1", "true", "True")
+
+# Largest-first open models if fallbacks are enabled
 OPEN_LLM_CANDIDATES: List[str] = [
-    "Qwen/Qwen2.5-72B-Instruct",           # large & open weights
-    "meta-llama/Meta-Llama-3.1-70B-Instruct",  # large (requires strong GPUs)
-    "mistralai/Mixtral-8x22B-Instruct-v0.1",   # MoE, open
-    "NousResearch/Meta-Llama-3.1-8B-Instruct", # smaller safety net
-    "mistralai/Mistral-7B-Instruct-v0.3"
+    "Qwen/Qwen2.5-72B-Instruct",
+    "meta-llama/Meta-Llama-3.1-70B-Instruct",
+    "mistralai/Mixtral-8x22B-Instruct-v0.1",
+    "NousResearch/Meta-Llama-3.1-8B-Instruct",
+    "mistralai/Mistral-7B-Instruct-v0.3",
 ]
 
 LOCAL_MAX_NEW_TOKENS = int(os.getenv("LOCAL_MAX_NEW_TOKENS", "1200"))
 
-# App knobs
+# ===== App Knobs =====
 MODEL_SETTINGS: Dict[str, Any] = {
     "temperature": float(os.getenv("TEMP", "0.3")),
     "top_p": float(os.getenv("TOP_P", "0.9")),
@@ -47,10 +54,8 @@ HEALTHCARE_SYSTEM_PROMPT = (
     "Do not invent numbers."
 )
 
-# Feature flags
 USE_SCENARIO_ENGINE = os.getenv("USE_SCENARIO_ENGINE", "1") not in ("0", "false", "False")
 
-# Storage
 DATA_DIR = os.getenv("DATA_DIR", "./data")
 RAG_INDEX_DIR = os.getenv("RAG_INDEX_DIR", "./rag_index")
 PERSIST_CONTENT = False