Update app.py

app.py

@@ -13,6 +13,7 @@ Notes:
  - This version includes a robust regex-based extractor that finds the outermost {...} block
    in the LLM output, extracts numeric values for the required keys, and always returns
    numeric defaults (no NaN) so frontends will not receive null/None for numeric fields.
+ - This variant logs raw LLM output and the parsed JSON using Python logging.
 """
 
 import io
@@ -41,6 +42,7 @@ except Exception:
 
 # Configure logging
 logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger("elderly_healthwatch")
 
 # Configuration for remote VLM and LLM spaces (change to your target Space names)
 GRADIO_VLM_SPACE = os.getenv("GRADIO_SPACE", "developer0hye/Qwen3-VL-8B-Instruct")
@@ -286,10 +288,10 @@ def run_vlm_and_get_features(face_path: str, eye_path: str, prompt: Optional[str
     message = {"text": prompt, "files": [handle_file(face_path), handle_file(eye_path)]}
 
     try:
-        logging.info("Calling VLM Space %s", GRADIO_VLM_SPACE)
+        logger.info("Calling VLM Space %s", GRADIO_VLM_SPACE)
         result = client.predict(message=message, history=[], api_name="/chat_fn")
     except Exception as e:
-        logging.exception("VLM call failed")
+        logger.exception("VLM call failed")
         raise RuntimeError(f"VLM call failed: {e}")
 
     if not result:
@@ -385,7 +387,7 @@ def run_llm_on_vlm(vlm_features_or_raw: Any,
     input_payload_str = instruction  # we feed only the instruction (which contains the VLM output)
 
     try:
-        logging.info("Calling LLM Space %s with strict JSON-only prompt", LLM_GRADIO_SPACE)
+        logger.info("Calling LLM Space %s with strict JSON-only prompt", LLM_GRADIO_SPACE)
         result = client.predict(
             input_data=input_payload_str,
             max_new_tokens=float(max_new_tokens),
@@ -400,7 +402,7 @@ def run_llm_on_vlm(vlm_features_or_raw: Any,
             api_name="/chat"
         )
     except Exception as e:
-        logging.exception("LLM call failed")
+        logger.exception("LLM call failed")
         raise RuntimeError(f"LLM call failed: {e}")
 
     # Normalize result to string
@@ -412,20 +414,30 @@ def run_llm_on_vlm(vlm_features_or_raw: Any,
     if not text_out or len(text_out.strip()) == 0:
         raise RuntimeError("LLM returned empty response")
 
+    # LOG raw output for debugging / auditing
+    logger.info("LLM raw output:\n%s", text_out)
+
     # Use regex-based extraction (robust)
     try:
         parsed = extract_json_via_regex(text_out)
     except Exception as e:
-        logging.exception("Regex JSON extraction failed")
+        logger.exception("Regex JSON extraction failed")
         # As a last fallback, attempt naive JSON parsing; if that fails, raise with raw output
         try:
             parsed = json.loads(text_out)
         except Exception:
+            # include raw output in the exception text so logs contain it
             raise ValueError(f"Failed to extract JSON from LLM output: {e}\nRaw Output:\n{text_out}")
 
     if not isinstance(parsed, dict):
         raise ValueError("Parsed LLM output is not a JSON object/dict")
 
+    # LOG parsed JSON (pretty-printed)
+    try:
+        logger.info("LLM parsed JSON:\n%s", json.dumps(parsed, indent=2, ensure_ascii=False))
+    except Exception:
+        logger.info("LLM parsed JSON (raw dict): %s", str(parsed))
+
     # Final safety clamps (already ensured by extractor, but keep defensive checks)
     def safe_prob(val):
         try:
@@ -679,10 +691,10 @@ async def process_screening(screening_id: str):
     """
     try:
         if screening_id not in screenings_db:
-            logging.error("[process_screening] screening %s not found", screening_id)
+            logger.error("[process_screening] screening %s not found", screening_id)
             return
         screenings_db[screening_id]["status"] = "processing"
-        logging.info("[process_screening] Starting %s", screening_id)
+        logger.info("[process_screening] Starting %s", screening_id)
 
         entry = screenings_db[screening_id]
         face_path = entry.get("face_image_path")
@@ -768,7 +780,7 @@ async def process_screening(screening_id: str):
                 "vlm_raw": vlm_raw
             })
         except Exception as e:
-            logging.exception("VLM feature extraction failed")
+            logger.exception("VLM feature extraction failed")
             screenings_db[screening_id].setdefault("ai_results", {})
             screenings_db[screening_id]["ai_results"].update({"vlm_error": str(e)})
             vlm_features = None
@@ -798,7 +810,7 @@ async def process_screening(screening_id: str):
             screenings_db[screening_id].setdefault("ai_results", {})
             screenings_db[screening_id]["ai_results"].update({"structured_risk": structured_risk})
         except Exception as e:
-            logging.exception("LLM processing failed")
+            logger.exception("LLM processing failed")
             screenings_db[screening_id].setdefault("ai_results", {})
             screenings_db[screening_id]["ai_results"].update({"llm_error": str(e)})
             structured_risk = {
@@ -850,14 +862,14 @@ async def process_screening(screening_id: str):
             "recommendations": recommendations
         })
 
-        logging.info("[process_screening] Completed %s", screening_id)
+        logger.info("[process_screening] Completed %s", screening_id)
     except Exception as e:
         traceback.print_exc()
         if screening_id in screenings_db:
             screenings_db[screening_id]["status"] = "failed"
             screenings_db[screening_id]["error"] = str(e)
         else:
-            logging.error("[process_screening] Failed for unknown screening %s: %s", screening_id, str(e))
+            logger.error("[process_screening] Failed for unknown screening %s: %s", screening_id, str(e))
 
 # -----------------------
 # Run server (for local debugging)