Update app.py

app.py

@@ -2,8 +2,8 @@ import gradio as gr
 import torch
 import json
 import os
-import fitz  # PyMuPDF für PDFs
-from transformers import AutoModelForCausalLM, AutoTokenizer
+import fitz
+from transformers import pipeline
 from huggingface_hub import hf_hub_download, HfApi
 from duckduckgo_search import DDGS
 from sentence_transformers import SentenceTransformer, util
@@ -13,10 +13,9 @@ DATASET_REPO = "Stemini/isaac-memory-db"
 HF_TOKEN = os.getenv("HF_TOKEN")
 api = HfApi()
 
-# Modelle laden
-model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForCausalLM.from_pretrained(model_name)
+# Schnelles Modell für CPU (Gradio 6 kompatibel)
+model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+pipe = pipeline("text-generation", model=model_id, torch_dtype=torch.bfloat16, device_map="cpu")
 search_model = SentenceTransformer('all-MiniLM-L6-v2')
 
 # --- FUNKTIONEN ---
@@ -31,101 +30,61 @@ def update_memory(new_memory):
     with open("memory.json", "w") as f: json.dump(new_memory, f)
     api.upload_file(path_or_fileobj="memory.json", path_in_repo="memory.json", repo_id=DATASET_REPO, repo_type="dataset", token=HF_TOKEN)
 
-def web_search(query):
-    try:
-        with DDGS() as ddgs:
-            results = [r['body'] for r in ddgs.text(query, max_results=2)]
-            return " ".join(results)
-    except: return "Keine Websuche möglich."
-
 def find_best_context(query, memory):
-    if not memory or len(memory) == 0: return "Kein Vorwissen."
+    if not memory: return "Kein Vorwissen."
     passages = list(memory.values())
     query_emb = search_model.encode(query, convert_to_tensor=True)
     passage_embs = search_model.encode(passages, convert_to_tensor=True)
     hits = util.semantic_search(query_emb, passage_embs, top_k=1)
     return passages[hits[0][0]['corpus_id']]
 
-def process_file(file):
-    if file is None: return "Keine Datei ausgewählt."
-    text = ""
-    if file.name.endswith(".pdf"):
-        doc = fitz.open(file.name)
-        for page in doc: text += page.get_text()
-    else:
-        with open(file.name, "r", encoding="utf-8") as f: text = f.read()
-    
-    memory = get_memory()
-    memory[str(len(memory))] = f"Dokument ({os.path.basename(file.name)}): {text[:1500]}"
-    update_memory(memory)
-    return f"Erfolg: '{os.path.basename(file.name)}' wurde integriert."
-
 def chat_logic(message, history):
     memory = get_memory()
-    
-    # 1. Korrektur-Befehl
-    if message.lower().startswith("korrektur:"):
-        info = message.replace("korrektur:", "").strip()
-        memory[str(len(memory))] = f"Manuelle Korrektur: {info}"
-        update_memory(memory)
-        history.append((message, "🧬 W-Vector Update: Wissen permanent verankert."))
-        return "", history
-
-    # 2. RAG & Delta I
     context = find_best_context(message, memory)
-    input_text = f"<|system|>\nKontext: {context}</s>\n<|user|>\n{message}</s>\n<|assistant|>\n"
-    inputs = tokenizer(input_text, return_tensors="pt")
     
-    with torch.no_grad():
-        outputs = model(**inputs, labels=inputs["input_ids"])
-        delta_i = outputs.loss.item()
-
-    # 3. Autonome Websuche
-    search_triggered = False
-    if delta_i > 2.0:
-        search_triggered = True
-        web_info = web_search(message)
-        memory[str(len(memory))] = f"Web-Wissen: {web_info[:300]}"
-        update_memory(memory)
-        input_text = f"<|system|>\nWeb-Daten: {web_info[:300]}</s>\n<|user|>\n{message}</s>\n<|assistant|>\n"
-        inputs = tokenizer(input_text, return_tensors="pt")
-
-    # 4. Generierung
-    gen_ids = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
-    answer = tokenizer.decode(gen_ids[0], skip_special_tokens=True).split("<|assistant|>\n")[-1]
+    # Prompt-Struktur
+    prompt = f"<|system|>\nKontext: {context}</s>\n<|user|>\n{message}</s>\n<|assistant|>\n"
+    
+    # Generierung (max 50 Tokens für Speed auf CPU)
+    outputs = pipe(prompt, max_new_tokens=50, do_sample=True, temperature=0.7)
+    answer = outputs[0]["generated_text"].split("<|assistant|>\n")[-1]
+    
+    # GRADIO 6 FORMAT: Liste aus Dictionaries
+    history.append({"role": "user", "content": message})
+    history.append({"role": "assistant", "content": answer})
     
-    # 5. Output für Gradio (Tupel-Format für Stabilität)
-    full_response = f"{answer}\n\n---\n📊 ΔI: {delta_i:.4f} | {'🌐 Web' if search_triggered else '🧠 Intern'}"
-    history.append((message, full_response))
     return "", history
 
-# --- GRADIO INTERFACE ---
+def process_file(file):
+    if file is None: return "Keine Datei."
+    text = ""
+    doc = fitz.open(file.name)
+    for page in doc: text += page.get_text()
+    memory = get_memory()
+    memory[str(len(memory))] = f"Doc: {text[:500]}"
+    update_memory(memory)
+    return "Integriert."
+
+# --- INTERFACE (GRADIO 6 OPTIMIERT) ---
 
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# Isaac: Evolution 2.0 Master-Node")
+    gr.Markdown("# Isaac: Evolution 2.0")
     
-    chatbot = gr.Chatbot(height=450, label="Isaac Chat")
+    # WICHTIG: type="messages" ist in Gradio 6 Pflicht für Dictionaries
+    chatbot = gr.Chatbot(height=450, label="Isaac Chat", type="messages")
     
     with gr.Row():
-        msg = gr.Textbox(
-            label="Deine Nachricht", 
-            placeholder="Hier tippen...",
-            show_label=True,
-            scale=4
-        )
-        submit_btn = gr.Button("Senden", variant="primary", scale=1)
+        msg = gr.Textbox(label="Deine Nachricht", placeholder="Tippen...", scale=4)
+        btn = gr.Button("Senden", variant="primary", scale=1)
 
     with gr.Row():
-        with gr.Column():
-            file_upload = gr.File(label="PDF/TXT Upload", file_types=[".pdf", ".txt"])
-        with gr.Column():
-            upload_status = gr.Textbox(label="Status", interactive=False)
+        upl = gr.File(label="PDF/TXT Upload", scale=1)
+        stat = gr.Textbox(label="Status", interactive=False, scale=1)
 
-    # Event-Handling für Enter und Button
+    # Event-Handling
     msg.submit(chat_logic, [msg, chatbot], [msg, chatbot])
-    submit_btn.click(chat_logic, [msg, chatbot], [msg, chatbot])
-    
-    file_upload.upload(process_file, file_upload, upload_status)
+    btn.click(chat_logic, [msg, chatbot], [msg, chatbot])
+    upl.upload(process_file, upl, stat)
 
 if __name__ == "__main__":
     demo.launch()