app.py

import gradio as gr
from transformers import pipeline
from langdetect import detect
from huggingface_hub import InferenceClient
import pandas as pd
import os

HF_TOKEN = os.getenv("HF_TOKEN")

# Fonction pour appeler l'API Zephyr
def call_zephyr_api(messages, hf_token=HF_TOKEN):
    client = InferenceClient("HuggingFaceH4/zephyr-7b-beta", token=hf_token)
    try:
        response = client.chat_completion(messages, max_tokens=300)
        return response.choices[0].message.content
    except Exception as e:
        raise gr.Error(f"❌ Erreur d'appel API Hugging Face : {str(e)}")

# Chargement du modèle de sentiment
classifier = pipeline("sentiment-analysis", model="mrm8488/distilroberta-finetuned-financial-news-sentiment-analysis")

# Modèles de traduction
translator_to_en = pipeline("translation", model="Helsinki-NLP/opus-mt-mul-en")
translator_to_fr = pipeline("translation", model="Helsinki-NLP/opus-mt-en-fr")

# Fonction pour suggérer le meilleur modèle
def suggest_model(text):
    word_count = len(text.split())
    if word_count < 50:
        return "Rapide"
    elif word_count <= 200:
        return "Équilibré"
    else:
        return "Précis"

# Fonction d'analyse
def full_analysis(text, mode, detail_mode, count, history):
    if not text:
        return "Entrez une phrase.", "", "", 0, history, None

    try:
        lang = detect(text)
    except:
        lang = "unknown"

    if lang != "en":
        text = translator_to_en(text, max_length=512)[0]['translation_text']

    result = classifier(text)[0]
    sentiment_output = f"Sentiment : {result['label']} (Score: {result['score']:.2f})"

    messages = [
        {"role": "system", "content": "You are a professional financial analyst AI."},
        {"role": "user", "content": f"Analyze the following financial news carefully:\n\"{text}\"\n\nThe detected sentiment for this news is: {result['label'].lower()}.\n\nNow, explain why the sentiment is {result['label'].lower()} using a logical, fact-based explanation.\nBase your reasoning only on the given news text.\nDo not repeat the news text or the prompt.\nRespond only with your financial analysis in one clear paragraph.\nWrite in a clear and professional tone."}
    ]
    explanation_en = call_zephyr_api(messages)

    explanation_en = call_zephyr_api(messages)
    explanation_fr = translator_to_fr(explanation_en, max_length=512)[0]['translation_text']

    count += 1
    history.append({
        "Texte": text,
        "Sentiment": result['label'],
        "Score": f"{result['score']:.2f}",
        "Explication_EN": explanation_en,
        "Explication_FR": explanation_fr
    })

    return sentiment_output, explanation_en, explanation_fr, count, history

# Fonction pour télécharger historique CSV
def download_history(history):
    if not history:
        return None
    df = pd.DataFrame(history)
    file_path = "/tmp/analysis_history.csv"
    df.to_csv(file_path, index=False)
    return file_path

# Interface Gradio
def launch_app():
    with gr.Blocks(theme=gr.themes.Base(), css="body {background-color: #0D1117; color: white;} .gr-button {background-color: #161B22; border: 1px solid #30363D;}") as iface:
        gr.Markdown("# 📈 Analyse Financière Premium + Explication IA", elem_id="title")
        gr.Markdown("Entrez une actualité financière. L'IA analyse et explique en anglais/français. Choisissez votre mode d'explication.")

        count = gr.State(0)
        history = gr.State([])

        with gr.Row():
            input_text = gr.Textbox(lines=4, placeholder="Entrez une actualité ici...", label="Texte à analyser")

        with gr.Row():
            mode_selector = gr.Dropdown(
                choices=["Rapide", "Équilibré", "Précis"],
                value="Équilibré",
                label="Mode recommandé selon la taille"
            )
            detail_mode_selector = gr.Dropdown(
                choices=["Normal", "Expert"],
                value="Normal",
                label="Niveau de détail"
            )

        analyze_btn = gr.Button("Analyser")
        reset_graph_btn = gr.Button("Reset Graphique")
        download_btn = gr.Button("Télécharger CSV")

        with gr.Row():
            sentiment_output = gr.Textbox(label="Résultat du Sentiment")

        with gr.Row():
            with gr.Column():
                explanation_output_en = gr.Textbox(label="Explication en Anglais")
            with gr.Column():
                explanation_output_fr = gr.Textbox(label="Explication en Français")

        download_file = gr.File(label="Fichier CSV")

        input_text.change(lambda t: gr.update(value=suggest_model(t)), inputs=[input_text], outputs=[mode_selector])

        analyze_btn.click(
            full_analysis,
            inputs=[input_text, mode_selector, detail_mode_selector, count, history],
            outputs=[sentiment_output, explanation_output_en, explanation_output_fr, count, history]
        )

        download_btn.click(
            download_history,
            inputs=[history],
            outputs=[download_file]
        )

    iface.launch()

if __name__ == "__main__":
    launch_app()