Update app.py

app.py

@@ -10,6 +10,53 @@ from autogluon.tabular import TabularPredictor
 MODEL_REPO_ID = "rlogh/cheese-texture-autogluon-classifier"
 DATASET_ID    = "aslan-ng/cheese-tabular"
 
+import os, json
+import urllib.request
+
+USE_HF_LOCAL = os.getenv("USE_HF", "0").lower() in {"1", "true", "yes"}
+USE_HF_API   = os.getenv("USE_HF_API", "0").lower() in {"1", "true", "yes"}
+HF_MODEL     = os.getenv("HF_MODEL", "google/flan-t5-small")
+HF_TOKEN     = os.getenv("HUGGINGFACEHUB_API_TOKEN")
+
+def _hf_inference_api(prompt: str) -> str:
+    """Call HF Inference API for text2text; avoids local downloads."""
+    url = f"https://api-inference.huggingface.co/models/{HF_MODEL}"
+    req = urllib.request.Request(
+        url,
+        data=json.dumps({"inputs": prompt}).encode("utf-8"),
+        headers={"Authorization": f"Bearer {HF_TOKEN}", "Content-Type": "application/json"},
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=60) as resp:
+        data = json.loads(resp.read().decode("utf-8"))
+    # HF returns either a list of dicts with 'generated_text' or a dict with 'generated_text'
+    if isinstance(data, list) and data and "generated_text" in data[0]:
+        return data[0]["generated_text"].strip()
+    if isinstance(data, dict) and "generated_text" in data:
+        return data["generated_text"].strip()
+    # Fallback if provider returns a different schema
+    return str(data).strip()
+
+def explain(structured: dict, engine: str = "auto") -> str:
+    prompt = build_prompt(structured)
+    # 1) Remote API (preferred for Spaces)
+    if USE_HF_API and HF_TOKEN:
+        try:
+            return _hf_inference_api(prompt)
+        except Exception as e:
+            return explain_template(structured) + f"\n\n_Explanation engine fell back to template: {e}_"
+    # 2) Local transformers (optional – see Fix B)
+    if USE_HF_LOCAL:
+        try:
+            from transformers import pipeline
+            pipe = pipeline("text2text-generation", model=HF_MODEL)
+            out = pipe(prompt, max_new_tokens=180, do_sample=False)
+            return out[0]["generated_text"].strip()
+        except Exception as e:
+            return explain_template(structured) + f"\n\n_Explanation engine fell back to template: {e}_"
+    # 3) Deterministic template
+    return explain_template(structured)
+
 def _safe_concat_splits(ds):
     frames = []
     for split in ds.keys():