Update settings.py

settings.py

@@ -1,25 +1,19 @@
 import os
 from typing import Dict, Any, List
 
-# ===== Residency / Deployment =====
+# Residency (informational; enforce via your infra)
 CANADA_RESIDENCY_REQUIRED = True
-CLOUD_REGION = os.getenv("CLOUD_REGION", "ca")  # e.g., "ca" or "us"
+CLOUD_REGION = os.getenv("CLOUD_REGION", "ca")
 
-# ===== Cohere API =====
+# ---- 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 models as needed)
+COHERE_API_URL = os.getenv("COHERE_API_URL", "")  # optional private base
+COHERE_TIMEOUT_S = float(os.getenv("COHERE_TIMEOUT_S", "60"))
 COHERE_MODEL_PRIMARY = os.getenv("COHERE_MODEL_PRIMARY", "command-r")
 COHERE_EMBED_MODEL = os.getenv("COHERE_EMBED_MODEL", "embed-english-v3.0")
 
-# ===== Fallbacks (OPEN MODELS ONLY) =====
-# By request, we default to Cohere-only. Set to "1" to allow open-model fallbacks.
+# ---- Open-model fallback (opt-in) ----
 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",
     "meta-llama/Meta-Llama-3.1-70B-Instruct",
@@ -27,15 +21,15 @@ OPEN_LLM_CANDIDATES: List[str] = [
     "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")),
     "repetition_penalty": float(os.getenv("REP_PEN", "1.12")),
     "max_new_tokens": int(os.getenv("MAX_NEW_TOKENS", "1500")),
+    "timeout_s": COHERE_TIMEOUT_S,
 }
 
 HEALTHCARE_SETTINGS = {
@@ -50,13 +44,18 @@ HEALTHCARE_SETTINGS = {
 GENERAL_CONVERSATION_PROMPT = "You are a helpful, concise assistant."
 HEALTHCARE_SYSTEM_PROMPT = (
     "You are a Canadian healthcare operations copilot. "
-    "You must follow the scenario tasks exactly; calculations are done deterministically. "
-    "Do not invent numbers."
+    "Follow the scenario tasks exactly. Calculations are done deterministically in code; "
+    "do not invent numbers."
 )
 
 USE_SCENARIO_ENGINE = os.getenv("USE_SCENARIO_ENGINE", "1") not in ("0", "false", "False")
 
+# Debug flags
+DEBUG_PLAN = os.getenv("DEBUG_PLAN", "1") in ("1", "true", "True")
+
+# Storage/local
 DATA_DIR = os.getenv("DATA_DIR", "./data")
 RAG_INDEX_DIR = os.getenv("RAG_INDEX_DIR", "./rag_index")
 PERSIST_CONTENT = False
 SNAPSHOT_PATH = os.getenv("SNAPSHOT_PATH", "./snapshots")
+