Update app.py

app.py

@@ -14,6 +14,9 @@ CHANGELOG (corrections applied):
   7. Prompt includes "topics_discussed" field so the report only covers relevant sections
   8. Resilience: retry logic (up to 3 attempts), robust JSON parser, raw response
      surfaced on failure for debugging
+  9. Hallucination guard: prompt instructs model to return not_a_planning_conversation
+     flag when audio does not contain advance care planning content; validation rejects
+     such responses before generating a report.
 """
 
 import os
@@ -51,6 +54,34 @@ logger = logging.getLogger(__name__)
 EXTRACTION_PROMPT = """
 You are analyzing a recorded conversation about advance care planning and end-of-life wishes.
 Listen to the ENTIRE audio carefully and extract ALL relevant information.
+
+IMPORTANT - NON-RELEVANT AUDIO DETECTION:
+Before extracting any planning data, first determine whether this audio actually
+contains an advance care planning conversation. The audio MUST contain a real
+discussion about at least one of these topics: health care wishes, financial
+planning, funeral/memorial preferences, or values and legacy.
+
+If the audio is:
+- Silence, noise, music, or unintelligible speech
+- A conversation about unrelated topics (e.g. casual chat, a lecture, a podcast
+  not about advance care planning)
+- Too short or too unclear to extract meaningful planning information
+- Random test audio or microphone checks
+
+Then return ONLY this JSON and nothing else:
+```json
+{
+  "not_a_planning_conversation": true,
+  "reason": "Brief explanation of what the audio actually contains"
+}
+```
+
+Do NOT invent, fabricate, or hallucinate planning data. If you cannot clearly hear
+a real advance care planning discussion, you MUST return the above JSON.
+
+Only proceed with the full extraction below if you are confident the audio contains
+a genuine advance care planning conversation.
+
 CRITICAL INSTRUCTIONS FOR SINGLE-SELECT FIELDS:
 - You MUST select exactly ONE option for each single-select field
 - Use the EXACT string values specified (copy them exactly)
@@ -311,6 +342,42 @@ def parse_json_response(response_text: str) -> dict | None:
     return None
 
 
+def _is_valid_planning_data(data: dict) -> tuple[bool, str]:
+    """Check whether parsed data represents a genuine planning conversation.
+
+    Returns (is_valid, reason).  When is_valid is False, reason contains a
+    user-facing message explaining why the audio was rejected.
+    """
+    if not data:
+        return False, "No data could be extracted from the audio."
+
+    # Explicit flag set by the model when the audio is not relevant
+    if data.get("not_a_planning_conversation"):
+        reason = data.get("reason", "The audio does not appear to contain an advance care planning conversation.")
+        return False, (
+            "This audio does not contain an advance care planning conversation. "
+            f"({reason}) Please record or upload a conversation that discusses "
+            "health care wishes, financial planning, or funeral preferences."
+        )
+
+    # Secondary heuristic: if none of the core sections are present, the
+    # model may have returned something unexpected.
+    has_any_section = any(
+        key in data for key in (
+            "health_care_wishes", "financial_planning", "funeral_plans",
+            "participant", "topics_discussed",
+        )
+    )
+    if not has_any_section:
+        return False, (
+            "The audio could not be matched to a planning conversation. "
+            "Please make sure the recording contains a discussion about "
+            "health care wishes, financial planning, or funeral preferences."
+        )
+
+    return True, ""
+
+
 def normalize_value(value, valid_options, default=None):
     """Normalize a value to match one of the valid options."""
     if value is None:
@@ -956,6 +1023,15 @@ def process_audio(audio_file):
                 raw_response,
             )
 
+        # ---- Hallucination guard ----
+        is_valid, rejection_reason = _is_valid_planning_data(data)
+        if not is_valid:
+            return (
+                None,
+                rejection_reason,
+                json.dumps(data, indent=2),
+            )
+
         # Normalize data
         data = normalize_data(data)
 
@@ -1108,4 +1184,4 @@ if __name__ == "__main__":
     if HAS_GRADIO:
         demo.launch(theme=custom_theme)
     else:
-        print("Gradio not installed. Core logic is available for import.") 
+        print("Gradio not installed. Core logic is available for import.")