🐛 Fix: improved model output post-processing to avoid empty answers

app.py

@@ -151,23 +151,25 @@ def ask_question(question):
 
     context = "\n".join([doc.page_content for doc in docs])
     prompt = generate_prompt(context, question)
-    response = reasoning_pipeline(prompt)[0]['generated_text']
+    output = reasoning_pipeline(prompt)[0]['generated_text'].strip()
 
-    # Clean out prompt leakage
-    for token in ["Context:", "Question:", "Instructions:", "Use structured academic language"]:
-        response = response.replace(token, "").strip()
+    # Clean unwanted leftovers
+    for marker in ["Context:", "Question:", "Instructions:", "Use structured academic language"]:
+        output = output.replace(marker, "").strip()
 
-    # Remove leading/trailing junk
-    if "Answer:" in response:
-        response = response.split("Answer:")[-1].strip()
-    if "." in response:
-        response = response.rsplit(".", 1)[0] + "."
+    # Remove leading "Answer:" if present
+    if output.lower().startswith("answer:"):
+        output = output[len("answer:"):].strip()
 
-    # ❗ Fallback if answer is empty or nonsense
-    if len(response.strip()) < 10:
-        return "❌ The model could not generate a meaningful answer based on the provided context."
+    # Ensure proper sentence ending
+    if "." in output:
+        output = output.rsplit(".", 1)[0] + "."
 
-    return post_process_output(response.strip(), question)
+    # Fallback if answer too short or generic
+    if len(output) < 10 or output.lower() in ["", ".", "use structured academic language.", "use structured academic language"]:
+        return "❌ The model could not generate a meaningful answer from the provided material."
+
+    return post_process_output(output, question)
 
 
 # Gradio UI