app.py

import gradio as gr
import os
import re
import requests
import numpy as np
import torch
from sklearn.neighbors import NearestNeighbors
from sentence_transformers import SentenceTransformer
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

# --- CONFIGURATION ---
HF_TOKEN = os.getenv("HF_TOKEN", "").strip()
HF_MODEL = "HuggingFaceH4/zephyr-7b-beta"  # Change this if needed
HF_API_URL = f"https://api-inference.huggingface.co/models/{HF_MODEL}"
headers = {"Authorization": f"Bearer {HF_TOKEN}"}

FILES = ["main1.txt", "main2.txt", "main3.txt", "main4.txt", "main5.txt", "main6.txt"]
EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"

EMBEDDING_CACHE_FILE = "embeddings.npy"
CHUNKS_CACHE_FILE = "chunks.npy"

# --- FUNCTIONS ---

def test_model_connection():
    try:
        print("🔍 Testing Hugging Face model availability...")
        test_response = requests.get(HF_API_URL, headers=headers, timeout=10)
        print("Status Code:", test_response.status_code)
        print("Response JSON:", test_response.json())
    except Exception as e:
        print("❌ Connection Test Failed:", e)

def load_text_files(file_list):
    knowledge = ""
    for file_name in file_list:
        try:
            with open(file_name, "r", encoding="utf-8") as f:
                knowledge += "\n" + f.read()
        except Exception as e:
            print(f"Error reading {file_name}: {e}")
    return knowledge.strip()

def chunk_text(text, max_chunk_length=500):
    sentences = re.split(r'(?<=[.!?])\s+', text)
    chunks = []
    current_chunk = ""
    for sentence in sentences:
        if len(current_chunk) + len(sentence) <= max_chunk_length:
            current_chunk += " " + sentence
        else:
            chunks.append(current_chunk.strip())
            current_chunk = sentence
    if current_chunk:
        chunks.append(current_chunk.strip())
    return chunks

def embed_texts(texts):
    return model.encode(texts, convert_to_numpy=True, normalize_embeddings=True)

def save_cache(embeddings, chunks):
    np.save(EMBEDDING_CACHE_FILE, embeddings)
    np.save(CHUNKS_CACHE_FILE, np.array(chunks))

def load_cache():
    if os.path.exists(EMBEDDING_CACHE_FILE) and os.path.exists(CHUNKS_CACHE_FILE):
        embeddings = np.load(EMBEDDING_CACHE_FILE, allow_pickle=True)
        chunks = np.load(CHUNKS_CACHE_FILE, allow_pickle=True).tolist()
        print("✅ Loaded cached embeddings and chunks.")
        return embeddings, chunks
    return None, None

def retrieve_chunks(query, top_k=5):
    query_embedding = embed_texts([query])
    distances, indices = nn_model.kneighbors(query_embedding, n_neighbors=top_k)
    return [chunks[i] for i in indices[0]]

def build_prompt(question):
    relevant_chunks = retrieve_chunks(question)
    context = "\n".join(relevant_chunks)

    system_instruction = """You are an AI-supported financial expert. Your role is to answer questions strictly within the context of the university lecture "Financial Markets" (Universität Duisburg-Essen).
Important instructions:
1. Base your answers primarily on the provided lecture excerpts ("lecture_slides").
2. If an answer is not directly covered by the lecture content, you may elaborate — but **only if you are absolutely certain**. Avoid making up information.
3. If you are unsure, reply politely: 
   "Entschuldigung. Leider kenne ich die Antwort auf diese Frage nicht."
4. If a formula is relevant, show the **exact formula** and explain it in **simple terms**.
5. Do not give vague or speculative answers — it's better to skip a question than guess.
6. **Never generate your own questions. Only respond to the given question.**
7. **Always respond in German.**
8. Make your answers clear, fact-based, and well-structured.
"""

    prompt = f"""{system_instruction}
Vorlesungsinhalte:
{context}
--- Ende der Vorlesungsinhalte ---
Frage des Nutzers (bitte nur diese beantworten): {question}
Antwort:"""
    return prompt

def respond(message, history):
    try:
        prompt = build_prompt(message)
        payload = {
            "inputs": prompt,
            "parameters": {
                "temperature": 0.2,
                "max_new_tokens": 400,
                "stop": ["Frage:", "Question:", "User:", "Frage des Nutzers"]
            },
        }

        response = requests.post(HF_API_URL, headers=headers, json=payload, timeout=60)
        response.raise_for_status()
        output = response.json()

        if isinstance(output, list) and "generated_text" in output[0]:
            generated_text = output[0]["generated_text"]
            answer = generated_text[len(prompt):].strip()
        else:
            print("❗️HF API returned unexpected format:", output)
            answer = "❌ Modell hat keine gültige Antwort geliefert. Bitte später erneut versuchen."

    except Exception as e:
        print("API Error:", e)
        try:
            print("Raw HF response:", response.text)
        except:
            pass
        answer = "❌ Error contacting the model. Please check your token, timeout, or model availability."

    if history is None:
        history = []

    history.append({"role": "assistant", "content": answer})
    return answer

# --- INIT SECTION ---

print("🔄 Initializing embedding model...")
model = SentenceTransformer(EMBEDDING_MODEL)

chunk_embeddings, chunks = load_cache()

if chunk_embeddings is None or chunks is None:
    print("🛠 No cache found. Processing text...")
    knowledge_base = load_text_files(FILES)
    chunks = chunk_text(knowledge_base)
    chunk_embeddings = embed_texts(chunks)
    save_cache(chunk_embeddings, chunks)
    print("✅ Embeddings and chunks cached.")

nn_model = NearestNeighbors(metric="cosine")
nn_model.fit(chunk_embeddings)

# --- GRADIO INTERFACE ---

demo = gr.ChatInterface(
    fn=respond,
    title="📚 RAG Chatbot Finanzmärkte",
    description="Stelle Fragen basierend auf den hochgeladenen Vorlesungstexten.",
    chatbot=gr.Chatbot(type="messages"),
)

if __name__ == "__main__":
    test_model_connection()
    demo.launch(debug=True)