fix: connect Peg to local AI runtime

server/app/coach.py

@@ -48,7 +48,8 @@ async def chat_with_peg(data: CoachChatIn, cfg: Settings) -> CoachChatOut:
             response = await client.post(
                 f"{cfg.ai_base_url.rstrip('/')}/chat/completions",
                 headers=headers,
-                json={"model": cfg.ai_model, "messages": messages, "temperature": 0.4, "max_tokens": 350},
+                # Keep answers useful but short so a local CPU model responds reliably.
+                json={"model": cfg.ai_model, "messages": messages, "temperature": 0.4, "max_tokens": 160},
             )
             response.raise_for_status()
             content = response.json()["choices"][0]["message"]["content"].strip()
@@ -60,3 +61,16 @@ async def chat_with_peg(data: CoachChatIn, cfg: Settings) -> CoachChatOut:
             return CoachChatOut(message=content, source="model")
     except (httpx.HTTPError, KeyError, IndexError, TypeError, ValueError):
         return CoachChatOut(message="Je suis momentanément indisponible. En attendant, choisis une petite action sans risque : boire un verre d’eau, marcher quelques minutes si tu te sens bien, ou noter ton prochain repas. Pour une question médicale, contacte un professionnel de santé.", source="fallback")
+
+
+async def peg_provider_health(cfg: Settings) -> dict[str, str | bool]:
+    """Check the configured AI provider without leaking credentials or prompts."""
+    base_url = cfg.ai_base_url.rstrip("/")
+    headers = {"Authorization": f"Bearer {cfg.ai_api_key}"} if cfg.ai_api_key else {}
+    try:
+        async with httpx.AsyncClient(timeout=min(cfg.ai_timeout_seconds, 5.0)) as client:
+            response = await client.get(f"{base_url}/models", headers=headers)
+            response.raise_for_status()
+        return {"status": "ok", "reachable": True, "model": cfg.ai_model}
+    except httpx.HTTPError:
+        return {"status": "unavailable", "reachable": False, "model": cfg.ai_model}

server/app/config.py

@@ -11,8 +11,9 @@ class Settings(BaseSettings):
     environment: str = "development"
     ai_base_url: str = "http://localhost:11434/v1"
     ai_api_key: str = ""
-    ai_model: str = "llama3.2:3b"
-    ai_timeout_seconds: float = 20.0
+    ai_model: str = "qwen2.5:1.5b"
+    # Local models may need extra time for their first response while loading.
+    ai_timeout_seconds: float = 90.0
     model_config = SettingsConfigDict(env_file=".env", extra="ignore")
 
     @property

server/app/main.py

@@ -10,7 +10,7 @@ from .config import get_settings
 from .db import Base, engine, get_db
 from .models import Entry, Profile, User
 from .schemas import CoachChatIn, CoachChatOut, EntryIn, EntryOut, ProfileIn, ProfileOut, Register, Token
-from .coach import chat_with_peg
+from .coach import chat_with_peg, peg_provider_health
 from .security import current_user, hash_password, make_token, verify_password
 
 cfg = get_settings()
@@ -31,6 +31,11 @@ def health():
     return {"status": "ok", "medical_notice": "MetaBloom does not diagnose or replace medical care."}
 
 
+@app.get("/health/ai")
+async def ai_health():
+    return await peg_provider_health(cfg)
+
+
 @app.post("/coach/chat", response_model=CoachChatOut)
 async def coach_chat(data: CoachChatIn, user: User = Depends(current_user)):
     return await chat_with_peg(data, cfg)