app.py

import streamlit as st
import pandas as pd
import torch
import requests
import os
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login

HF_TOKEN = os.getenv("Allie") or "<your_token_here>"
if HF_TOKEN:
    login(HF_TOKEN)

# Define model map
model_map = {
    "InvestLM": {"id": "yixuantt/InvestLM-mistral-AWQ", "local": False},
    "FinLLaMA": {"id": "us4/fin-llama3.1-8b", "local": False},
    "FinanceConnect": {"id": "ceadar-ie/FinanceConnect-13B", "local": True},
    "Sujet-Finance": {"id": "sujet-ai/Sujet-Finance-8B-v0.1", "local": True},
    "FinGPT (LoRA)": {"id": "FinGPT/fingpt-mt_llama2-7b_lora", "local": True}  # Placeholder, special handling below
}

# Load question list
@st.cache_data
def load_questions():
    df = pd.read_csv("questions.csv")
    return df["Question"].dropna().tolist()

# Load local models
@st.cache_resource
def load_local_model(model_id):
    tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=HF_TOKEN)
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.float32,
        device_map="auto",
        use_auth_token=HF_TOKEN
    )
    return model, tokenizer

# Prompt template
PROMPT_TEMPLATE = (
    "You are FinGPT, a highly knowledgeable and reliable financial assistant.\n"
    "Explain the following finance/tax/controlling question clearly, including formulas, examples, and reasons why it matters.\n"
    "\n"
    "Question: {question}\n"
    "Answer:"
)

# Local generation
def query_local_model(model_id, prompt):
    model, tokenizer = load_local_model(model_id)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=400,
        temperature=0.7,
        top_p=0.9,
        top_k=40,
        repetition_penalty=1.2,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Remote HF inference
def query_remote_model(model_id, prompt):
    headers = {"Authorization": f"Bearer {HF_TOKEN}"}
    payload = {"inputs": prompt, "parameters": {"max_new_tokens": 400}}
    response = requests.post(
        f"https://api-inference.huggingface.co/models/{model_id}",
        headers=headers,
        json=payload
    )
    result = response.json()
    return result[0]["generated_text"] if isinstance(result, list) else result.get("generated_text", "ERROR")

# Route to appropriate model
def query_model(model_entry, question):
    prompt = PROMPT_TEMPLATE.format(question=question)
    if model_entry["id"] == "FinGPT/fingpt-mt_llama2-7b_lora":
        return "⚠️ FinGPT (LoRA) integration requires manual loading with PEFT and is not available via HF API."
    elif model_entry["local"]:
        return query_local_model(model_entry["id"], prompt)
    else:
        return query_remote_model(model_entry["id"], prompt)

# === UI ===
st.set_page_config(page_title="Finanzmodell Tester", layout="centered")
st.title("📊 Finanzmodell Vergleichs-Interface")

questions = load_questions()
question_choice = st.selectbox("Wähle eine Frage", questions)
model_choice = st.selectbox("Wähle ein Modell", list(model_map.keys()))

if st.button("Antwort generieren"):
    with st.spinner("Antwort wird generiert..."):
        model_entry = model_map[model_choice]
        try:
            answer = query_model(model_entry, question_choice)
        except Exception as e:
            answer = f"[Fehler: {str(e)}]"
        st.text_area("💬 Antwort des Modells:", value=answer, height=400, disabled=True)