Update app.py

app.py

@@ -64,23 +64,22 @@ def get_conv_names():
     with lock:
         return list(conversations.keys())
 
-# FORMAT ULTRA-MINIMALISTE POUR MAXIMUM DE VITESSE
 def build_conversation_prompt(history, new_message):
-    """Format ultra-optimisé pour la vitesse maximale"""
+    """Format de prompt optimisé pour Alpaca"""
+    prompt = ""
     
-    # System prompt très court, seulement au début
-    if not history:
-        prompt = "Tu es Alisia, assistante IA. Réponds en français.\n\n"
-    else:
-        prompt = ""
+    # System prompt seulement au début
+    if not any(any(conv) for conv in conversations.values()):
+        prompt += """Tu es Alisia, une assistante IA utile et compétente. Réponds de manière précise et concise.
+
+"""
     
-    # Historique formaté de façon minimaliste
-    for i, (user_msg, assistant_msg) in enumerate(history):
-        prompt += f"Q: {user_msg}\nA: {assistant_msg}\n\n"
+    # Historique de conversation
+    for user_msg, assistant_msg in history:
+        prompt += f"### Instruction:\n{user_msg}\n\n### Response:\n{assistant_msg}\n\n"
     
     # Nouveau message
-    prompt += f"Q: {new_message}\nA:"
-    
+    prompt += f"### Instruction:\n{new_message}\n\n### Response:\n"
     return prompt
 
 def send_message_stream(user_message, displayed_history, current_chat_name):
@@ -99,31 +98,26 @@ def send_message_stream(user_message, displayed_history, current_chat_name):
     local_hist.append((str(user_message), ""))
     yield local_hist, ""
 
-    # Construction ULTRA-RAPIDE du prompt
     formatted_prompt = build_conversation_prompt(local_hist[:-1], str(user_message))
     partial = ""
     
-    # PARAMÈTRES DE RÉACTIVITÉ MAXIMALE
+    # PARAMÈTRES DE RÉACTIVITÉ HYBRIDE
     last_update = time.time()
     token_count = 0
-    min_tokens = 1      # Mise à jour après chaque token
-    max_delay = 0.05    # Seulement 50ms entre les updates!
+    min_tokens = 2      # Minimum de tokens avant update
+    max_delay = 0.12    # Maximum 120ms entre updates
 
     try:
-        # Paramètres de génération ULTRA-OPTIMISÉS
         stream = llm.create_completion(
             prompt=formatted_prompt,
             stream=True,
-            max_tokens=512,   # Réduit au maximum pour la vitesse
+            max_tokens=1024,
             temperature=0.7,
             top_p=0.9,
             repeat_penalty=1.1,
-            stop=["Q:", "A:", "\n\n", "<|endoftext|>", "###", "Instruction:", "Response:"],
-            top_k=40,
-            min_p=0.05        # Ajouté pour la vitesse
+            stop=["### Instruction:", "### Response:", "\n\n", "<|endoftext|>"]
         )
 
-        buffer = ""
         for chunk in stream:
             if stop_generation.is_set():
                 break
@@ -131,34 +125,27 @@ def send_message_stream(user_message, displayed_history, current_chat_name):
             if "choices" in chunk and chunk["choices"]:
                 token = chunk["choices"][0].get("text", "")
                 if token:
-                    buffer += token
+                    partial += token
                     token_count += 1
                     
-                    # STRATÉGIE ULTRA-RAPIDE - mise à jour immédiate
-                    current_time = time.time()
-                    time_since_update = current_time - last_update
-                    
+                    # STRATÉGIE DE RÉACTIVITÉ HYBRIDE
                     should_update = (
                         token_count >= min_tokens or 
-                        time_since_update >= max_delay or
-                        token in ["\n", ".", "!", "?", ","]
+                        time.time() - last_update > max_delay or
+                        token in [".", "!", "?", "\n", ",", ";", ":"]
                     )
                     
-                    if should_update and buffer.strip():
-                        cleaned = clean_output(buffer)
-                        if cleaned:  # Only update if we have meaningful content
-                            current_response = local_hist[-1][1] + cleaned
-                            local_hist[-1] = (str(user_message), current_response)
-                            yield local_hist, ""
-                            buffer = ""  # Clear buffer after updating
-                        last_update = current_time
+                    if should_update:
+                        cleaned = clean_output(partial)
+                        local_hist[-1] = (str(user_message), cleaned)
+                        yield local_hist, ""
+                        last_update = time.time()
                         token_count = 0
 
-        # Final update with any remaining content
-        if buffer.strip():
-            cleaned = clean_output(buffer)
-            current_response = local_hist[-1][1] + cleaned
-            local_hist[-1] = (str(user_message), current_response)
+        # DERNIER FLUSH - Garantit que tout est affiché
+        if partial:
+            cleaned = clean_output(partial)
+            local_hist[-1] = (str(user_message), cleaned)
             yield local_hist, ""
 
     except Exception as e:
@@ -382,7 +369,7 @@ with gr.Blocks(css=css, title="Alisia Chat - Ultra Rapide", theme=gr.themes.Soft
         menu_btn = gr.Button("☰", elem_classes="hamburger")
         gr.Markdown("### 💬 Alisia <span class='alisia-badge'>AI Assistant</span>", elem_id="title")
         gr.HTML("<div style='flex:1'></div>")
-        gr.Markdown("<small style='color:#94a3b8'>Mode Turbo Activé</small>")
+        gr.Markdown("<small style='color:#94a3b8'>llama.cpp optimisé</small>")
 
     with gr.Row():
         with gr.Column(scale=1, visible=True, elem_id="leftcol") as left_column:
@@ -403,12 +390,12 @@ with gr.Blocks(css=css, title="Alisia Chat - Ultra Rapide", theme=gr.themes.Soft
                     "🗑️ Effacer chat",
                     elem_classes="clear-btn"
                 )
-                gr.Markdown("## 🚀 Mode Turbo", elem_classes="conversation-header")
+                gr.Markdown("## 🚀 Mode Ultra-Rapide", elem_classes="conversation-header")
                 gr.Markdown("""
                 <div style="color: #94a3b8; font-size: 14px;">
-                ✅ Format minimaliste<br>
-                ✅ Réactivité 50ms<br>
-                ✅ Optimisation max
+                ✅ Streaming hybride<br>
+                ✅ Réactivité 120ms<br>
+                ✅ Optimisé llama.cpp
                 </div>
                 """, elem_classes="conversation-subheader")
 
@@ -508,9 +495,8 @@ with gr.Blocks(css=css, title="Alisia Chat - Ultra Rapide", theme=gr.themes.Soft
 # LANCEMENT
 # -------------------------
 if __name__ == "__main__":
-    print("🚀 Lancement du mode TURBO...")
-    print("⏱️  Réactivité maximale (50ms)")
-    print("🎯 Format ultra-minimaliste")
+    print("🚀 Lancement de l'interface ultra-réactive...")
+    print("⏱️  Mode streaming hybride activé (120ms)")
     demo.launch(
         share=True,
         server_name="0.0.0.0",