Update backend/chat_endpoint.py

backend/chat_endpoint.py

@@ -9,33 +9,86 @@ def _headers():
     tok = os.getenv("HF_API_TOKEN")
     return {"Authorization": f"Bearer {tok}","Content-Type":"application/json"} if tok else {}
 
+
 def chat(user_message: str, mode: str = "patient"):
     url = active_chat_endpoint()
-    if not _headers():
+    headers = _headers()
+    if not headers:
         return "⚠ Add HF_API_TOKEN in Settings → Secrets."
-    system = AI_GYNO_PERSONA_V2 + ("\nPrefer plain-language for patients." if mode=='patient' else "\nProvide differentials, workup, and red flags.")
+
+    system = AI_GYNO_PERSONA_V2 + (
+        "\nUse simple, reassuring language." if mode=="patient"
+        else "\nUse clinical language, differentials, and red flags."
+    )
+
     payload = {
         "inputs": [
-            {"role":"system","content": system},
-            {"role":"user","content": user_message}
+            {"role": "system", "content": system},
+            {"role": "user", "content": user_message}
         ],
-        "parameters": {"max_new_tokens": 400, "temperature": 0.2, "return_full_text": False}
+        "parameters": {
+            "max_new_tokens": 400,
+            "temperature": 0.2,
+            "return_full_text": False
+        }
     }
-    for attempt in range(1, RETRIES_DEFAULT+1):
-        try:
-            r = requests.post(url, headers=_headers(), json=payload, timeout=REQUEST_TIMEOUT_SECONDS_DEFAULT)
-            try:
-                data = r.json()
-            except Exception:
-                txt = r.text
-                if "loading" in txt.lower():
-                    time.sleep(BACKOFF_SECONDS_DEFAULT * attempt); continue
-                return f"⚠ Non-JSON response:\n\n{txt[:1500]}"
-            if isinstance(data, list) and data and "generated_text" in data[0]:
-                return data[0]["generated_text"]
-            if isinstance(data, dict) and "generated_text" in data:
-                return data["generated_text"]
-            return "⚠ Unexpected response:\n" + json.dumps(data)[:1200]
-        except Exception as e:
-            time.sleep(BACKOFF_SECONDS_DEFAULT * attempt)
-    return "❌ Endpoint unavailable after retries."
+
+    try:
+        r = requests.post(url, headers=headers, json=payload, timeout=60)
+    except Exception as e:
+        return f"⚠ Network error: {str(e)}"
+
+    # --- SAFER PARSER ---
+    txt = r.text
+
+    # HTML → clear signal model is loading or wrong
+    if "<html" in txt.lower():
+        if "loading" in txt.lower():
+            return "⏳ Model is loading on HuggingFace… try again in 20–30 sec."
+        return f"⚠ Endpoint returned HTML instead of JSON:\n\n{txt[:500]}"
+
+    # Try normal JSON
+    try:
+        data = r.json()
+    except:
+        return f"⚠ Non-JSON content returned:\n\n{txt[:500]}"
+
+    # Standard HF generation
+    if isinstance(data, list) and len(data) and "generated_text" in data[0]:
+        return data[0]["generated_text"]
+    if isinstance(data, dict) and "generated_text" in data:
+        return data["generated_text"]
+
+    return f"⚠ Unexpected JSON:\n{json.dumps(data)[:800]}"
+
+
+# def chat(user_message: str, mode: str = "patient"):
+#     url = active_chat_endpoint()
+#     if not _headers():
+#         return "⚠ Add HF_API_TOKEN in Settings → Secrets."
+#     system = AI_GYNO_PERSONA_V2 + ("\nPrefer plain-language for patients." if mode=='patient' else "\nProvide differentials, workup, and red flags.")
+#     payload = {
+#         "inputs": [
+#             {"role":"system","content": system},
+#             {"role":"user","content": user_message}
+#         ],
+#         "parameters": {"max_new_tokens": 400, "temperature": 0.2, "return_full_text": False}
+#     }
+#     for attempt in range(1, RETRIES_DEFAULT+1):
+#         try:
+#             r = requests.post(url, headers=_headers(), json=payload, timeout=REQUEST_TIMEOUT_SECONDS_DEFAULT)
+#             try:
+#                 data = r.json()
+#             except Exception:
+#                 txt = r.text
+#                 if "loading" in txt.lower():
+#                     time.sleep(BACKOFF_SECONDS_DEFAULT * attempt); continue
+#                 return f"⚠ Non-JSON response:\n\n{txt[:1500]}"
+#             if isinstance(data, list) and data and "generated_text" in data[0]:
+#                 return data[0]["generated_text"]
+#             if isinstance(data, dict) and "generated_text" in data:
+#                 return data["generated_text"]
+#             return "⚠ Unexpected response:\n" + json.dumps(data)[:1200]
+#         except Exception as e:
+#             time.sleep(BACKOFF_SECONDS_DEFAULT * attempt)
+#     return "❌ Endpoint unavailable after retries."