Update result messages

app.py

@@ -1,7 +1,7 @@
 import gradio as gr
 from dotenv import load_dotenv
 
-from utils import get_evaluated_component, get_evaluated_component_adjective, to_title_case, get_prompt_from_test_case, to_snake_case
+from utils import get_evaluated_component, get_evaluated_component_adjective, get_result_description, to_title_case, get_prompt_from_test_case, to_snake_case
 load_dotenv()
 import json
 from model import generate_text
@@ -49,12 +49,10 @@ def on_submit(criteria, context, user_message, assistant_message, state):
         'user_message': user_message,
         'assistant_message': assistant_message,
     }, state['selected_sub_catalog'])
-    evaluated_component = get_evaluated_component(state['selected_sub_catalog'], state['selected_criteria_name'])
-    evaluated_component_adjective = get_evaluated_component_adjective(state['selected_sub_catalog'], state['selected_criteria_name'])
     logger.debug(f"Starting evaluation for subcatelog {state['selected_sub_catalog']} and criteria name {state['selected_criteria_name']}")
     result_label = generate_text(prompt)['assessment'] # Yes or No
     # result_str = f"Is the {evaluated_component} message harmful based on the risk definition: {result_label}"
-    html_str = f"<p>Is the {evaluated_component} {evaluated_component_adjective}: <strong>{result_label}</strong></p>"
+    html_str = f"<p>{get_result_description(state['selected_sub_catalog'], state['selected_criteria_name'])} <strong>{result_label}</strong></p>"
     # html_str = f"Is the {evaluated_component} {evaluated_component_adjective}: {result_label}"
     return [gr.update(value=html_str), gr.update(visible=True), gr.update(visible=True)]
 

model.py

@@ -58,7 +58,7 @@ def generate_text(prompt):
     mock_model_call = os.getenv('MOCK_MODEL_CALL') == 'true'
     if mock_model_call:
         logger.debug('Returning mocked model result.')
-        sleep(2)
+        sleep(1)
         return {'assessment': 'Yes', 'certainty': 0.97}
     else:
         start = time()

utils.py

@@ -20,6 +20,21 @@ def get_prompt_template(test_case, sub_catalog_name):
 def get_prompt_from_test_case(test_case, sub_catalog_name):
     return assessment_prompt(Template(get_prompt_template(test_case, sub_catalog_name)).render(**test_case))
 
+def get_result_description(sub_catalog_name, criteria_name):
+    evaluated_component = get_evaluated_component(sub_catalog_name, criteria_name)
+    messages = {
+        "general_harm": f"Is the {evaluated_component} message harmful by common-sense?",
+        "bias": f"Is the {evaluated_component} message prejudiced based on identity or characteristics?",
+        "jailbreak": f"Does the user message try to manipulate the AI to generate harmful, undesired, or inappropriate content?",
+        "violence": f"Does the {evaluated_component} message contain content promoting physical, mental, or sexual harm?",
+        "profanity": f"Does the {evaluated_component} message include offensive language or insults?",
+        "unethical_behavior": f"Does the {evaluated_component} message include actions that violate moral or legal standards?",
+        "answer_relevance": f"Does the assistant response fail to address or properly answer the user question?",
+        "context_relevance": f"Is the retrieved context irrelevant to the user question or does not address their needs?",
+        "groundedness": f"Does the assistant response include claims or facts not supported by or contradicted by the provided context?"
+    }
+    return messages[criteria_name]
+
 def get_evaluated_component(sub_catalog_name, criteria_name):
     if sub_catalog_name == 'harmful_content_in_user_message':
         component = "user"
@@ -32,7 +47,7 @@ def get_evaluated_component(sub_catalog_name, criteria_name):
             component = "assistant"
         elif criteria_name == "answer_relevance":
             component = "assistant"
-    return f"'{to_title_case(component)} message'"
+    return component
 
 def get_evaluated_component_adjective(sub_catalog_name, criteria_name):
     if criteria_name == 'context_relevance' or criteria_name == 'answer_relevance':