Upload app.py

app.py

@@ -10,10 +10,13 @@ from huggingface_hub import InferenceClient
 # ────────────────────────────────────────────────────────────────────────────────
 # CONFIG
 
-# You can hard-code the model here, or override with MODEL_ID as a Space variable.
-# Use a free, public chat model that doesn't need a novita api_key
-MODEL_ID = os.environ.get("MODEL_ID", "TinyLlama/TinyLlama-1.1B-Chat-v1.0")
-HF_TOKEN = os.environ.get("HF_TOKEN")  # can stay empty
+# Default to a serverless text-generation model. You can override this in Space
+# settings by defining a MODEL_ID variable if you want to experiment.
+MODEL_ID = os.environ.get("MODEL_ID", "google/gemma-2-2b-it")
+
+# Personal access token from your Hugging Face account (Space secret).
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
 
 
 REPO_PATH = "/data/questions.json"  # where we store generated questions
@@ -335,10 +338,19 @@ def try_parse_json(text: str) -> Optional[Dict[str, Any]]:
 
 # 🔧 SIMPLIFIED, ROBUST MODEL CALL (no secrets required)
 def model_call(prompt: str) -> str:
+    """
+    Call Hugging Face Inference API using a single text-generation call.
+    """
     if not MODEL_ID:
-        raise RuntimeError("MODEL_ID env var is not set and no default is provided.")
+        raise RuntimeError("MODEL_ID env var is empty. Set it or use the default.")
 
-    client = InferenceClient(model=MODEL_ID, token=HF_TOKEN or None)
+    # HF_TOKEN must be defined as a Secret in the Space (read-only token is enough)
+    if not HF_TOKEN:
+        raise RuntimeError(
+            "HF_TOKEN is not set. Add a Hugging Face token as a Space secret named HF_TOKEN."
+        )
+
+    client = InferenceClient(model=MODEL_ID, token=HF_TOKEN)
 
     try:
         out = client.text_generation(
@@ -349,10 +361,11 @@ def model_call(prompt: str) -> str:
             return_full_text=False,
         )
     except Exception as e:
+        # Bubble up a clear error so ai_generate can surface it
         raise RuntimeError(f"Inference API error: {e}") from e
 
     if not out:
-        raise RuntimeError("Empty response from model.")
+        raise RuntimeError("Inference API returned an empty response.")
 
     return out.strip()
 
@@ -387,7 +400,12 @@ def normalize_output(
 
 
 def ai_generate(lang: str, category_key: str, variant: str) -> Dict[str, Any]:
+    """
+    Try to call the model. If anything fails or JSON is invalid,
+    fall back to shuffling the few-shots and include a safety_notes message.
+    """
     prompt = build_prompt(lang, category_key, variant)
+
     try:
         raw_text = model_call(prompt)
         parsed = try_parse_json(raw_text) if raw_text else None
@@ -395,29 +413,40 @@ def ai_generate(lang: str, category_key: str, variant: str) -> Dict[str, Any]:
         if parsed:
             return normalize_output(parsed, lang, category_key, variant)
 
-        # If model replied but JSON is bad, show that in safety_notes
+        # Model replied but not valid JSON
         few = FEWSHOTS[lang][category_key]
+        q_pool = few["questions"][:]
+        m_pool = few["micro_actions"][:]
+        random.shuffle(q_pool)
+        random.shuffle(m_pool)
         return {
             "category": category_key,
             "language": lang,
-            "questions": random.sample(few["questions"], 4),
-            "micro_actions": random.sample(few["micro_actions"], 2),
+            "questions": (q_pool + [""] * 4)[:4],
+            "micro_actions": (m_pool + [""] * 2)[:2],
             "tone": "fallback",
-            "safety_notes": f"Model replied but JSON parsing failed. raw_text starts with: {repr(raw_text[:160])}",
+            "safety_notes": (
+                "Model replied but JSON parsing failed. "
+                f"raw_text starts with: {repr(raw_text[:160])}"
+            ),
         }
+
     except Exception as e:
-        # Surface the error clearly
+        # Any HF / network / auth error ends up here
         few = FEWSHOTS[lang][category_key]
+        q_pool = few["questions"][:]
+        m_pool = few["micro_actions"][:]
+        random.shuffle(q_pool)
+        random.shuffle(m_pool)
         return {
             "category": category_key,
             "language": lang,
-            "questions": random.sample(few["questions"], 4),
-            "micro_actions": random.sample(few["micro_actions"], 2),
+            "questions": (q_pool + [""] * 4)[:4],
+            "micro_actions": (m_pool + [""] * 2)[:2],
             "tone": "error",
             "safety_notes": f"Model call error: {type(e).__name__}: {e}",
         }
 
-
 # ────────────────────────────────────────────────────────────────────────────────
 # MAIN LOGIC: REPO + SESSION "SEEN" QUESTIONS