pages/5_Conversation_Evaluation.py

import json

import pandas as pd
import streamlit as st
from util.evaluator import evaluator, write_evaluation_commentary
import os


# Predefined examples
examples = {
    'good': [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What causes rainbows to appear in the sky?"},
        {"role": "assistant",
         "content": "Rainbows appear when sunlight is refracted, dispersed, and reflected inside water droplets in the atmosphere, resulting in a spectrum of light appearing in the sky."},
        {"role": "user", "content": "That's interesting! Why does it create so many colors?"}
    ],
    'bad': [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What causes rainbows to appear in the sky?"},
        {"role": "assistant",
         "content": "Rainbows happen because light in the sky gets mixed up and sometimes shows colors when it's raining or when there is water around."},
        {"role": "user", "content": "That doesn't seem very clear."}
    ]
}


# Function to check password
def check_password():
    def password_entered():
        if password_input == os.getenv('PASSWORD'):
            st.session_state['password_correct'] = True
        else:
            st.error("Incorrect Password, please try again.")

    password_input = st.text_input("Enter Password:", type="password")
    submit_button = st.button("Submit", on_click=password_entered)

    if submit_button and not st.session_state.get('password_correct', False):
        st.error("Please enter a valid password to access the demo.")


# Title of the application
st.title('Single Evaluation of Conversations')

# Description of the application
st.sidebar.write("""
### Welcome to the Single Evaluation of Conversations Demo
This application allows you to evaluate the quality of conversations generated for various contexts using different language models. You can either use predefined examples or input your own conversations and contexts.
""")

# Explanation of principles
st.sidebar.write("""
### Explanation Principles
When evaluating conversations, consider the following principles mapped to user empowerment and regulatory compliance outcomes:

1. **Factually Correct**: The information should be accurate and relevant to empower users and meet external audit requirements.
2. **Useful**: Explanations should be clear and meaningful, helping users make informed decisions.
3. **Context Specific**: Explanations should be tailored to the context of use, enhancing their relevance and utility.
4. **User Specific**: Explanations should address the needs and preferences of the user, enabling better decision-making.
5. **Provide Pluralism**: Explanations should present diverse perspectives, allowing users to understand different viewpoints and make well-rounded decisions.
""")

# Check if password has been validated
if not st.session_state.get('password_correct', False):
    check_password()
else:
    st.sidebar.success("Password Verified. Proceed with the demo.")
    model_name = st.selectbox('Select a model:', ['gpt4-1106', 'gpt35-1106'])

    # User choice between predefined examples or their own input
    input_type = st.radio("Choose input type:", ('Use predefined example', 'Enter your own'))

    if input_type == 'Use predefined example':
        example_type = st.radio("Select an example type:", ('good', 'bad'))
        conversation = examples[example_type]
        context = "User role: general user\nUser background: interested in atmospheric phenomena\nSystem role: AI assistant providing educational information\nScenario: The user is asking about the scientific explanation of rainbows."
    else:
        conversation_input = st.text_area('Enter your conversation (JSON format):',
                                          '[{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Who won the world series in 2020?"}, {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."}]')
        context_input = st.text_area('Enter your context:',
                                     'User role: sports fan\nUser background: keen interest in baseball\nSystem role: AI assistant providing sports information\nScenario: The user is asking about historical sports events.')

        try:
            conversation = json.loads(conversation_input)
            context = context_input
        except json.JSONDecodeError:
            st.error("Invalid JSON format for conversation.")
            conversation = None
            context = None

    st.write('### Conversation')
    if conversation:
        for exchange in conversation:
            role = exchange['role'].capitalize()
            content = exchange['content']
            st.markdown(f"**{role}:** {content}")
    else:
        st.write('No conversation entered yet.')

    st.write('### Context')
    if context:
        st.write(context)
    else:
        st.write('No context entered yet.')

    if st.button('Evaluate Conversation'):
        if conversation and context:
            eval = evaluator(model_name)
            scores = eval.evaluate_conversation(conversation, context)
            st.write('### Scores')
            details = write_evaluation_commentary(scores)
            df = pd.DataFrame(details)
            st.write(df)

            data = {
                'Conversation': conversation,
                'Context': context,
                **{detail['Principle']: detail['Score'] for detail in details}
            }
            df = pd.DataFrame([data])

            # Convert DataFrame to CSV for download
            csv = df.to_csv(index=False)
            st.download_button(
                label="Download evaluation as CSV",
                data=csv,
                file_name='evaluation.csv',
                mime='text/csv',
            )
        else:
            st.error('Please enter both a conversation and a context to evaluate.')