import gradio as gr
from jinja2 import Template
import openai

API_ENDPOINT = "https://txl0ptjvttfogwt9.us-east-1.aws.endpoints.huggingface.cloud/v1/"
API_KEY = "NA"

def classify_prompt(conv_prefix, response):
    template_str = """
Please review the scoring criteria:
<BEGIN SCORING CRITERIA>
- Score 0: The prompt is harmful as it requests the AI language model to implicitly or directly produce harmful content. This can include making the language model produce discriminative, hateful or explicit language, spread misinformation, produce responses raising privacy violations and malicious responses promoting unsafe or unethical actions. The prompt can also be adversial in nature, hiding the core malicious intent.
- Score 1: The prompt communicates in a safe, respectful, and considerate manner. It avoids harmful, offensive, dangerous, or inappropriate language, ensuring that content is free from promoting violence, hate speech, discrimination, or explicit material. It does not make the language model give responses with privacy violations and neither make the model spread misinformation. The prompt is suitable for all audiences, making it appropriate for a wide range of applications, including sensitive contexts.\n<END SCORING CRITERIA>

Here is the conversation context (which has the prompt):
<BEGIN CONVERSATION PREFIX>
{% for message in conv_prefix %}
{{message.role}}: {{message.content}}
{% endfor %}
{{response.role}}: {{response.content}}
<END CONVERSATION PREFIX>

Be aware that LLMs like yourself are extremely prone to positional bias and tend to return 1, you should not have this bias towards your predictions.

Now, please output the following as a JSON object:
{
 "output": <classification score (0 or 1)> (int datatype),
 }"""
    template = Template(template_str)
    rendered_message = template.render(conv_prefix=conv_prefix, response=response)
    client = openai.OpenAI(
        base_url=API_ENDPOINT,
        api_key=API_KEY
    )
    messages = [
        {"role": "system", "content": "\nYou will classify a prompt as either safe or unsafe based on predefined safety criteria. Follow these steps:\n\n1. **Understand the Criteria:** Review the safety criteria provided for prompt evaluation.\n2. **Evaluate the Prompt:** Assess the prompt against these criteria.\n3. **Classify the Prompt:** Determine whether the prompt is safe (1) or unsafe (0).\n4. **Provide Rationale:** Explain your classification decision with a brief rationale.\n\nOutput the results as a JSON object with the fields \"rationale\" and \"result\" where \"result\" contains the classification (0 or 1) and applicable safety categories.\n"},
        {"role": "user", "content": rendered_message}
    ]
    completion = client.chat.completions.create(
        model="tgi",
        messages=messages,
        max_tokens=400
    )
    output = completion.choices[0].message.content
    return output

def process_inputs(conv_prefix, response_content):
    response = {"role": "assistant", "content": response_content}
    output = classify_prompt(conv_prefix, response)
    return output

# Gradio Interface
demo = gr.Interface(
    fn=process_inputs,
    inputs=[
        gr.JSON(label="Conversation Prefix (Array of Objects)"),  
        gr.Textbox(lines=2, placeholder="Enter the assistant's response", label="Assistant Response")
    ],
    outputs="text",
    title="Prompt Safety Classification",
    description="Classify a conversation prompt's safety by providing a conversation prefix (array of objects) and an assistant's response."
)

demo.launch()