app.py

import os
import gradio as gr
from huggingface_hub import InferenceClient

# Load Hugging Face token from secret
client = InferenceClient(
    provider="nscale",  # You can change to 'openrouter' or 'novita' if needed
    api_key=os.environ["HF_TOKEN"],
)

# Test prompt list
preset_prompts = [
    "I finally got the promotion, but I feel guilty because my best friend got laid off.",
    "Moving to a new city is exciting, but leaving my family breaks my heart.",
    "I passed the test, but my friend failed — and I don’t know how to feel.",
    "They applauded me on stage, but all I could think about was how lonely I felt.",
    "I’m happy for her, but I wish I had that too.",
]

# Core generation logic using chat completion
def call_llama(messages):
    try:
        completion = client.chat.completions.create(
            model="meta-llama/Llama-3.1-8B-Instruct",
            messages=messages,
        )
        return completion.choices[0].message.content.strip()
    except Exception as e:
        return f"⚠️ Error: {str(e)}"

# Emotion pipeline
def emotion_annotator(text):
    # Step 1: List candidate emotions
    msg1 = [
        {
            "role": "user",
            "content": f'List all possible emotions the person might be feeling in this sentence:\n"{text}"\nJust give comma-separated emotion names.'
        }
    ]
    candidates = call_llama(msg1)

    # Step 2: Choose most likely emotion with reason
    msg2 = [
        {
            "role": "user",
            "content": f'From these emotions: {candidates}, which is most likely the dominant one in the sentence "{text}"? Explain why briefly.\nFormat:\nMost likely emotion: <emotion>\nReason: <why>'
        }
    ]
    final = call_llama(msg2)

    return candidates, final

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## 🧠 Emotion Annotator (LLaMA 3.1 via Hugging Face Chat API)")
    gr.Markdown("Powered by `meta-llama/Llama-3.1-8B-Instruct`, served using the InferenceClient chat interface.")

    with gr.Row():
        text_input = gr.Textbox(label="✏️ Input Sentence", placeholder="e.g., I’m proud but I feel like I let them down.")
        dropdown = gr.Dropdown(preset_prompts, label="💬 Choose an example")
    run_button = gr.Button("Submit")

    with gr.Row():
        candidate_output = gr.Textbox(label="🧠 Candidate Emotions")
        final_output = gr.Textbox(label="🎯 Most Likely Emotion + Explanation")

    # Dropdown autofill
    dropdown.change(fn=lambda x: x, inputs=dropdown, outputs=text_input)
    run_button.click(fn=emotion_annotator, inputs=text_input, outputs=[candidate_output, final_output])

demo.launch()