vibe coding lead to a very lazy and wrong application of constrained decoding - simply cutting off the llm after a closing brace, bad bad bad - also I'm not sure why it was trying to deliver json at all, this plays to all of the weaknesses of gen ai, bad llm bad

app.py

@@ -106,14 +106,9 @@ def on_submit(symptoms_text, history):
     yield history, None, "\n".join(log)
 
     # Call LLM
-    # Use constrained decoding to enforce JSON-only output
-    response = llm.complete(prompt, stop=["}"]) # stop after closing brace
+    response = llm.complete(prompt)
     raw = getattr(response, 'text', str(response))
-    # Truncate extra content after the final JSON object
-    if not raw.strip().endswith('}'):
-        end_idx = raw.rfind('}')
-        if end_idx != -1:
-            raw = raw[:end_idx+1]
+
     msg = "📡 Raw LLM response received"
     log.append(msg)
     print(msg, flush=True)
@@ -121,23 +116,17 @@ def on_submit(symptoms_text, history):
 
     # Parse JSON
     cleaned_raw = re.sub(r",\s*([}\]])", r"\1", raw)
-    try:
-        parsed = json.loads(cleaned_raw)
-        msg = "✅ JSON parsed"
-    except Exception as e:
-        msg = f"❌ JSON parse error: {e}"
-        parsed = {"error": str(e), "raw": raw}
     log.append(msg)
     print(msg, flush=True)
-    yield history, parsed, "\n".join(log)
+    yield history, raw, "\n".join(log)
 
     # Final assistant message
-    assistant_msg = format_response_for_user(parsed)
+    assistant_msg = format_response_for_user(raw)
     history = history + [{"role": "assistant", "content": assistant_msg}]
     msg = "✅ Final response appended"
     log.append(msg)
     print(msg, flush=True)
-    yield history, parsed, "\n".join(log)
+    yield history, raw, "\n".join(log)
 
 # ========== Gradio UI ==========
 with gr.Blocks(theme="default") as demo: