Spaces:
Running
on
Zero
Running
on
Zero
Upload step03_chatbot.py with huggingface_hub
Browse files- step03_chatbot.py +1553 -0
step03_chatbot.py
ADDED
|
@@ -0,0 +1,1553 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Step 03 - Interface de chat RAG générique avec Gradio
|
| 4 |
+
Utilise les embeddings de Step 02 depuis Hugging Face Hub + Qwen3-4B-Instruct-2507 pour génération
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import os
|
| 8 |
+
import json
|
| 9 |
+
import numpy as np
|
| 10 |
+
import gradio as gr
|
| 11 |
+
from gradio import ChatMessage
|
| 12 |
+
from typing import List, Dict, Optional, Tuple
|
| 13 |
+
import time
|
| 14 |
+
import torch
|
| 15 |
+
import threading
|
| 16 |
+
import http.server
|
| 17 |
+
import socketserver
|
| 18 |
+
from pathlib import Path
|
| 19 |
+
from datetime import datetime
|
| 20 |
+
|
| 21 |
+
# ZeroGPU compatibility
|
| 22 |
+
try:
|
| 23 |
+
import spaces
|
| 24 |
+
ZEROGPU_AVAILABLE = True
|
| 25 |
+
print("🚀 ZeroGPU détecté - activation du support")
|
| 26 |
+
except ImportError:
|
| 27 |
+
ZEROGPU_AVAILABLE = False
|
| 28 |
+
# Fallback decorator for local usage
|
| 29 |
+
class MockSpaces:
|
| 30 |
+
@staticmethod
|
| 31 |
+
def GPU(duration=None):
|
| 32 |
+
def decorator(func):
|
| 33 |
+
return func
|
| 34 |
+
return decorator
|
| 35 |
+
spaces = MockSpaces()
|
| 36 |
+
|
| 37 |
+
def _check_dependencies():
|
| 38 |
+
"""Vérifie les dépendances nécessaires."""
|
| 39 |
+
missing = []
|
| 40 |
+
try:
|
| 41 |
+
import torch
|
| 42 |
+
except ImportError:
|
| 43 |
+
missing.append("torch")
|
| 44 |
+
|
| 45 |
+
try:
|
| 46 |
+
import numpy as np
|
| 47 |
+
except ImportError:
|
| 48 |
+
missing.append("numpy")
|
| 49 |
+
|
| 50 |
+
try:
|
| 51 |
+
from safetensors.torch import load_file
|
| 52 |
+
except ImportError:
|
| 53 |
+
missing.append("safetensors")
|
| 54 |
+
|
| 55 |
+
try:
|
| 56 |
+
from huggingface_hub import hf_hub_download
|
| 57 |
+
except ImportError:
|
| 58 |
+
missing.append("huggingface-hub")
|
| 59 |
+
|
| 60 |
+
try:
|
| 61 |
+
import faiss
|
| 62 |
+
except ImportError:
|
| 63 |
+
missing.append("faiss-cpu")
|
| 64 |
+
|
| 65 |
+
try:
|
| 66 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSequenceClassification
|
| 67 |
+
except ImportError:
|
| 68 |
+
missing.append("transformers")
|
| 69 |
+
|
| 70 |
+
try:
|
| 71 |
+
from sentence_transformers import SentenceTransformer
|
| 72 |
+
except ImportError:
|
| 73 |
+
missing.append("sentence-transformers")
|
| 74 |
+
|
| 75 |
+
if missing:
|
| 76 |
+
print(f"❌ Dépendances manquantes: {', '.join(missing)}")
|
| 77 |
+
print("📦 Installer avec: pip install " + " ".join(missing))
|
| 78 |
+
return False
|
| 79 |
+
return True
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
class Step03Config:
|
| 83 |
+
"""Gestionnaire de configuration Step 03 basé sur la sortie Step 02."""
|
| 84 |
+
|
| 85 |
+
def __init__(self, config_file: str = "step03_config.json"):
|
| 86 |
+
self.config_file = Path(config_file)
|
| 87 |
+
self.config = self.load_config()
|
| 88 |
+
|
| 89 |
+
def load_config(self) -> Dict:
|
| 90 |
+
"""Charge la configuration Step 03."""
|
| 91 |
+
if not self.config_file.exists():
|
| 92 |
+
raise FileNotFoundError(
|
| 93 |
+
f"❌ Configuration Step 03 non trouvée: {self.config_file}\n"
|
| 94 |
+
f"💡 Lancez d'abord: python step02_upload_embeddings.py"
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
try:
|
| 98 |
+
with open(self.config_file, 'r', encoding='utf-8') as f:
|
| 99 |
+
config = json.load(f)
|
| 100 |
+
|
| 101 |
+
# Vérification de la structure
|
| 102 |
+
if not config.get("step02_completed"):
|
| 103 |
+
raise ValueError("❌ Step 02 non complété selon la configuration")
|
| 104 |
+
|
| 105 |
+
required_keys = ["huggingface", "embeddings_info"]
|
| 106 |
+
for key in required_keys:
|
| 107 |
+
if key not in config:
|
| 108 |
+
raise ValueError(f"❌ Clé manquante dans configuration: {key}")
|
| 109 |
+
|
| 110 |
+
return config
|
| 111 |
+
|
| 112 |
+
except json.JSONDecodeError as e:
|
| 113 |
+
raise ValueError(f"❌ Configuration Step 03 malformée: {e}")
|
| 114 |
+
|
| 115 |
+
@property
|
| 116 |
+
def repo_id(self) -> str:
|
| 117 |
+
"""Repository Hugging Face ID."""
|
| 118 |
+
return self.config["huggingface"]["repo_id"]
|
| 119 |
+
|
| 120 |
+
@property
|
| 121 |
+
def dataset_name(self) -> str:
|
| 122 |
+
"""Nom du dataset."""
|
| 123 |
+
return self.config["huggingface"]["dataset_name"]
|
| 124 |
+
|
| 125 |
+
@property
|
| 126 |
+
def embeddings_file(self) -> str:
|
| 127 |
+
"""Nom du fichier SafeTensors."""
|
| 128 |
+
return self.config["huggingface"]["files"]["embeddings"]
|
| 129 |
+
|
| 130 |
+
@property
|
| 131 |
+
def metadata_file(self) -> str:
|
| 132 |
+
"""Nom du fichier métadonnées."""
|
| 133 |
+
return self.config["huggingface"]["files"]["metadata"]
|
| 134 |
+
|
| 135 |
+
@property
|
| 136 |
+
def total_vectors(self) -> int:
|
| 137 |
+
"""Nombre total de vecteurs."""
|
| 138 |
+
return self.config["embeddings_info"]["total_vectors"]
|
| 139 |
+
|
| 140 |
+
@property
|
| 141 |
+
def vector_dimension(self) -> int:
|
| 142 |
+
"""Dimension des vecteurs."""
|
| 143 |
+
return self.config["embeddings_info"]["vector_dimension"]
|
| 144 |
+
|
| 145 |
+
@property
|
| 146 |
+
def embedding_model(self) -> str:
|
| 147 |
+
"""Modèle d'embedding utilisé."""
|
| 148 |
+
return self.config["embeddings_info"]["embedding_model"]
|
| 149 |
+
|
| 150 |
+
|
| 151 |
+
class Qwen3Reranker:
|
| 152 |
+
"""
|
| 153 |
+
Reranker utilisant Qwen3-Reranker-4B pour améliorer la pertinence des résultats de recherche
|
| 154 |
+
"""
|
| 155 |
+
|
| 156 |
+
def __init__(self, model_name: str = "Qwen/Qwen3-Reranker-4B", use_flash_attention: bool = True):
|
| 157 |
+
"""
|
| 158 |
+
Initialise le reranker Qwen3
|
| 159 |
+
|
| 160 |
+
Args:
|
| 161 |
+
model_name: Nom du modèle HuggingFace à charger
|
| 162 |
+
use_flash_attention: Utiliser Flash Attention 2 si disponible (auto-d��sactivé sur Mac)
|
| 163 |
+
"""
|
| 164 |
+
self.model_name = model_name
|
| 165 |
+
self.use_flash_attention = use_flash_attention
|
| 166 |
+
|
| 167 |
+
# Détection de l'environnement
|
| 168 |
+
self.is_mps = torch.backends.mps.is_available()
|
| 169 |
+
self.is_cuda = torch.cuda.is_available()
|
| 170 |
+
self.is_cpu = not self.is_mps and not self.is_cuda
|
| 171 |
+
|
| 172 |
+
print(f"🔄 Chargement du reranker {model_name}...")
|
| 173 |
+
self._detect_platform()
|
| 174 |
+
self._load_model()
|
| 175 |
+
|
| 176 |
+
def _detect_platform(self):
|
| 177 |
+
"""Détecte la plateforme et ajuste les paramètres"""
|
| 178 |
+
if self.is_mps:
|
| 179 |
+
print(" - Plateforme: Mac MPS détecté")
|
| 180 |
+
self.use_flash_attention = False # Flash Attention non compatible MPS
|
| 181 |
+
self.batch_size = 1 # Traitement strictement individuel sur Mac
|
| 182 |
+
self.memory_cleanup_freq = 3 # Nettoyage mémoire fréquent
|
| 183 |
+
elif self.is_cuda:
|
| 184 |
+
print(f" - Plateforme: CUDA détecté ({torch.cuda.get_device_name()})")
|
| 185 |
+
self.batch_size = 1 # Garde traitement individuel pour stabilité
|
| 186 |
+
self.memory_cleanup_freq = 10 # Nettoyage moins fréquent
|
| 187 |
+
else:
|
| 188 |
+
print(" - Plateforme: CPU")
|
| 189 |
+
self.use_flash_attention = False
|
| 190 |
+
self.batch_size = 1
|
| 191 |
+
self.memory_cleanup_freq = 5
|
| 192 |
+
|
| 193 |
+
def _load_model(self):
|
| 194 |
+
"""Charge le modèle et le tokenizer"""
|
| 195 |
+
try:
|
| 196 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 197 |
+
|
| 198 |
+
# Chargement du tokenizer
|
| 199 |
+
print(" - Chargement du tokenizer...")
|
| 200 |
+
self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
|
| 201 |
+
|
| 202 |
+
# Configuration du modèle selon la plateforme
|
| 203 |
+
model_kwargs = self._get_model_config()
|
| 204 |
+
|
| 205 |
+
# Chargement du modèle
|
| 206 |
+
print(" - Chargement du modèle...")
|
| 207 |
+
self.model = AutoModelForSequenceClassification.from_pretrained(
|
| 208 |
+
self.model_name,
|
| 209 |
+
**model_kwargs
|
| 210 |
+
)
|
| 211 |
+
|
| 212 |
+
# Configuration du device
|
| 213 |
+
self._setup_device()
|
| 214 |
+
|
| 215 |
+
print(f"✅ Reranker chargé sur {self.device}")
|
| 216 |
+
print(f" - Flash Attention: {'✅' if self.use_flash_attention else '❌'}")
|
| 217 |
+
print(f" - Paramètres: {self.get_parameter_count():.1f}B")
|
| 218 |
+
|
| 219 |
+
except Exception as e:
|
| 220 |
+
print(f"❌ Erreur lors du chargement du reranker: {e}")
|
| 221 |
+
print("💡 Le reranking sera désactivé")
|
| 222 |
+
self.model = None
|
| 223 |
+
self.tokenizer = None
|
| 224 |
+
self.device = None
|
| 225 |
+
|
| 226 |
+
def _get_model_config(self) -> Dict:
|
| 227 |
+
"""Retourne la configuration du modèle selon la plateforme"""
|
| 228 |
+
config = {}
|
| 229 |
+
|
| 230 |
+
if self.is_mps:
|
| 231 |
+
# Configuration pour Mac MPS
|
| 232 |
+
config["torch_dtype"] = torch.float32 # MPS fonctionne mieux avec float32
|
| 233 |
+
config["device_map"] = None # device_map peut causer des problèmes avec MPS
|
| 234 |
+
elif self.is_cuda:
|
| 235 |
+
# Configuration pour CUDA
|
| 236 |
+
config["torch_dtype"] = torch.float16
|
| 237 |
+
if self.use_flash_attention:
|
| 238 |
+
try:
|
| 239 |
+
config["attn_implementation"] = "flash_attention_2"
|
| 240 |
+
print(" - Flash Attention 2 activée")
|
| 241 |
+
except Exception:
|
| 242 |
+
print(" - Flash Attention 2 non disponible, utilisation standard")
|
| 243 |
+
self.use_flash_attention = False
|
| 244 |
+
else:
|
| 245 |
+
config["device_map"] = "auto"
|
| 246 |
+
else:
|
| 247 |
+
# Configuration pour CPU
|
| 248 |
+
config["torch_dtype"] = torch.float32
|
| 249 |
+
config["device_map"] = "cpu"
|
| 250 |
+
|
| 251 |
+
return config
|
| 252 |
+
|
| 253 |
+
def _setup_device(self):
|
| 254 |
+
"""Configure le device pour le modèle"""
|
| 255 |
+
if self.is_mps:
|
| 256 |
+
self.device = torch.device("mps")
|
| 257 |
+
self.model = self.model.to(self.device)
|
| 258 |
+
elif self.is_cuda:
|
| 259 |
+
if hasattr(self.model, 'device'):
|
| 260 |
+
self.device = next(self.model.parameters()).device
|
| 261 |
+
else:
|
| 262 |
+
self.device = torch.device("cuda")
|
| 263 |
+
self.model = self.model.to(self.device)
|
| 264 |
+
else:
|
| 265 |
+
self.device = torch.device("cpu")
|
| 266 |
+
self.model = self.model.to(self.device)
|
| 267 |
+
|
| 268 |
+
def _format_pair(self, query: str, document: str, instruction: str = None) -> str:
|
| 269 |
+
"""
|
| 270 |
+
Formate une paire query-document pour le reranker
|
| 271 |
+
"""
|
| 272 |
+
if instruction:
|
| 273 |
+
return f"Instruction: {instruction}\nQuery: {query}\nDocument: {document}"
|
| 274 |
+
return f"Query: {query}\nDocument: {document}"
|
| 275 |
+
|
| 276 |
+
def _get_default_instruction(self) -> str:
|
| 277 |
+
"""Retourne l'instruction par défaut pour la documentation technique"""
|
| 278 |
+
return (
|
| 279 |
+
"Évaluez la pertinence de ce document technique "
|
| 280 |
+
"par rapport à la requête en considérant : terminologie technique, "
|
| 281 |
+
"spécifications, normes, procédures de mise en œuvre."
|
| 282 |
+
)
|
| 283 |
+
|
| 284 |
+
def _process_single_document(self, query: str, document: str, instruction: str) -> float:
|
| 285 |
+
"""
|
| 286 |
+
Traite un seul document et retourne son score de pertinence
|
| 287 |
+
"""
|
| 288 |
+
# Formatage de la paire
|
| 289 |
+
pair_text = self._format_pair(query, document, instruction)
|
| 290 |
+
|
| 291 |
+
# Tokenisation (pas de problème de padding avec un seul document)
|
| 292 |
+
inputs = self.tokenizer(
|
| 293 |
+
pair_text,
|
| 294 |
+
truncation=True,
|
| 295 |
+
max_length=512,
|
| 296 |
+
return_tensors="pt",
|
| 297 |
+
padding=False
|
| 298 |
+
)
|
| 299 |
+
|
| 300 |
+
# Déplacement vers le device
|
| 301 |
+
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
| 302 |
+
|
| 303 |
+
# Inférence
|
| 304 |
+
with torch.no_grad():
|
| 305 |
+
outputs = self.model(**inputs)
|
| 306 |
+
logits = outputs.logits
|
| 307 |
+
|
| 308 |
+
# Le modèle Qwen3-Reranker retourne des logits de forme [1, 2]
|
| 309 |
+
# pour classification binaire : [non-pertinent, pertinent]
|
| 310 |
+
probs = torch.nn.functional.softmax(logits, dim=1)
|
| 311 |
+
score = probs[0, 1].cpu().item() # Classe 1 = pertinent
|
| 312 |
+
|
| 313 |
+
return float(score)
|
| 314 |
+
|
| 315 |
+
def _cleanup_memory(self):
|
| 316 |
+
"""Nettoie la mémoire selon la plateforme"""
|
| 317 |
+
if self.is_mps:
|
| 318 |
+
if hasattr(torch.mps, 'empty_cache'):
|
| 319 |
+
torch.mps.empty_cache()
|
| 320 |
+
elif self.is_cuda:
|
| 321 |
+
torch.cuda.empty_cache()
|
| 322 |
+
|
| 323 |
+
import gc
|
| 324 |
+
gc.collect()
|
| 325 |
+
|
| 326 |
+
def rerank(self, query: str, documents: List[str], instruction: str = None) -> List[float]:
|
| 327 |
+
"""
|
| 328 |
+
Reranke une liste de documents par rapport à une requête
|
| 329 |
+
"""
|
| 330 |
+
if not documents:
|
| 331 |
+
return []
|
| 332 |
+
|
| 333 |
+
if self.model is None or self.tokenizer is None:
|
| 334 |
+
print(" - Reranker non disponible, scores neutres retournés")
|
| 335 |
+
return [0.5] * len(documents)
|
| 336 |
+
|
| 337 |
+
if instruction is None:
|
| 338 |
+
instruction = self._get_default_instruction()
|
| 339 |
+
|
| 340 |
+
print(f" - Reranking de {len(documents)} documents (traitement individuel)")
|
| 341 |
+
|
| 342 |
+
scores = []
|
| 343 |
+
successful_count = 0
|
| 344 |
+
|
| 345 |
+
for i, document in enumerate(documents):
|
| 346 |
+
try:
|
| 347 |
+
score = self._process_single_document(query, document, instruction)
|
| 348 |
+
score = max(0.0, min(1.0, score))
|
| 349 |
+
scores.append(score)
|
| 350 |
+
successful_count += 1
|
| 351 |
+
|
| 352 |
+
if (i + 1) % self.memory_cleanup_freq == 0:
|
| 353 |
+
self._cleanup_memory()
|
| 354 |
+
|
| 355 |
+
except Exception as doc_error:
|
| 356 |
+
print(f" ⚠️ Erreur document {i+1}: {doc_error}")
|
| 357 |
+
scores.append(0.5) # Score neutre en cas d'erreur
|
| 358 |
+
|
| 359 |
+
self._cleanup_memory()
|
| 360 |
+
|
| 361 |
+
print(f" ✅ Reranking terminé: {successful_count}/{len(documents)} documents traités")
|
| 362 |
+
|
| 363 |
+
if successful_count > 0:
|
| 364 |
+
valid_scores = [s for s in scores if s != 0.5]
|
| 365 |
+
if valid_scores:
|
| 366 |
+
top_scores = sorted(valid_scores, reverse=True)[:3]
|
| 367 |
+
print(f" 📈 Top 3 scores: {[f'{s:.3f}' for s in top_scores]}")
|
| 368 |
+
|
| 369 |
+
return scores
|
| 370 |
+
|
| 371 |
+
def get_parameter_count(self) -> float:
|
| 372 |
+
"""Retourne le nombre de paramètres du modèle en milliards"""
|
| 373 |
+
if self.model is None:
|
| 374 |
+
return 0.0
|
| 375 |
+
try:
|
| 376 |
+
return sum(p.numel() for p in self.model.parameters()) / 1e9
|
| 377 |
+
except:
|
| 378 |
+
return 0.0
|
| 379 |
+
|
| 380 |
+
def is_available(self) -> bool:
|
| 381 |
+
"""Vérifie si le reranker est disponible et fonctionnel"""
|
| 382 |
+
return self.model is not None and self.tokenizer is not None
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
class GenericRAGChatbot:
|
| 386 |
+
"""Chatbot RAG générique utilisant les embeddings de Step 02 et Qwen3-4B-Instruct pour la génération"""
|
| 387 |
+
|
| 388 |
+
def __init__(self,
|
| 389 |
+
generation_model: str = "Qwen/Qwen3-4B-Instruct-2507",
|
| 390 |
+
initial_k: int = 20,
|
| 391 |
+
final_k: int = 3,
|
| 392 |
+
use_flash_attention: bool = True,
|
| 393 |
+
use_reranker: bool = True):
|
| 394 |
+
"""
|
| 395 |
+
Initialise le système RAG générique
|
| 396 |
+
|
| 397 |
+
Args:
|
| 398 |
+
generation_model: Modèle Qwen3 pour la génération
|
| 399 |
+
initial_k: Nombre de candidats pour la recherche initiale
|
| 400 |
+
final_k: Nombre de documents finaux après reranking
|
| 401 |
+
use_flash_attention: Utiliser Flash Attention (désactivé automatiquement sur Mac)
|
| 402 |
+
use_reranker: Utiliser le reranking Qwen3
|
| 403 |
+
"""
|
| 404 |
+
self.generation_model_name = generation_model
|
| 405 |
+
self.initial_k = initial_k
|
| 406 |
+
self.final_k = final_k
|
| 407 |
+
self.use_flash_attention = use_flash_attention
|
| 408 |
+
self.use_reranker = use_reranker
|
| 409 |
+
|
| 410 |
+
# Détection de l'environnement (local + ZeroGPU)
|
| 411 |
+
self.is_zerogpu = ZEROGPU_AVAILABLE and os.getenv("SPACE_ID") is not None
|
| 412 |
+
self.is_mps = torch.backends.mps.is_available() and not self.is_zerogpu
|
| 413 |
+
self.is_cuda = torch.cuda.is_available()
|
| 414 |
+
|
| 415 |
+
# Configuration du device
|
| 416 |
+
if self.is_mps:
|
| 417 |
+
self.device = torch.device("mps")
|
| 418 |
+
elif self.is_cuda:
|
| 419 |
+
self.device = torch.device("cuda")
|
| 420 |
+
else:
|
| 421 |
+
self.device = torch.device("cpu")
|
| 422 |
+
|
| 423 |
+
if self.is_zerogpu:
|
| 424 |
+
print("🚀 Environnement ZeroGPU détecté - optimisations cloud")
|
| 425 |
+
self.use_flash_attention = True # ZeroGPU supporte Flash Attention
|
| 426 |
+
elif self.is_mps and use_flash_attention:
|
| 427 |
+
print("🍎 Mac avec MPS détecté - désactivation automatique de Flash Attention")
|
| 428 |
+
self.use_flash_attention = False
|
| 429 |
+
|
| 430 |
+
# Chargement des composants
|
| 431 |
+
self._load_step03_config()
|
| 432 |
+
self._load_embeddings_from_hf()
|
| 433 |
+
self._load_embedding_model()
|
| 434 |
+
self._load_reranker()
|
| 435 |
+
self._load_generation_model()
|
| 436 |
+
|
| 437 |
+
def _load_step03_config(self):
|
| 438 |
+
"""Charge la configuration Step 03"""
|
| 439 |
+
try:
|
| 440 |
+
self.config = Step03Config()
|
| 441 |
+
print(f"✅ Configuration Step 03 chargée")
|
| 442 |
+
print(f" 📦 Repository HF: {self.config.repo_id}")
|
| 443 |
+
print(f" 📊 Embeddings: {self.config.total_vectors:,} vecteurs")
|
| 444 |
+
print(f" 📏 Dimension: {self.config.vector_dimension}")
|
| 445 |
+
except Exception as e:
|
| 446 |
+
print(f"❌ Erreur de chargement de la configuration: {e}")
|
| 447 |
+
raise
|
| 448 |
+
|
| 449 |
+
def _load_embeddings_from_hf(self):
|
| 450 |
+
"""Télécharge et charge les embeddings depuis Hugging Face Hub"""
|
| 451 |
+
try:
|
| 452 |
+
from huggingface_hub import hf_hub_download
|
| 453 |
+
from safetensors.torch import load_file
|
| 454 |
+
import numpy as np
|
| 455 |
+
import faiss
|
| 456 |
+
|
| 457 |
+
print(f"🔄 Téléchargement des embeddings depuis {self.config.repo_id}...")
|
| 458 |
+
|
| 459 |
+
# Télécharger les fichiers (sans token pour les repos publics)
|
| 460 |
+
try:
|
| 461 |
+
embeddings_file = hf_hub_download(
|
| 462 |
+
repo_id=self.config.repo_id,
|
| 463 |
+
filename=self.config.embeddings_file,
|
| 464 |
+
repo_type="dataset",
|
| 465 |
+
token=None # Forcer l'accès sans token pour les repos publics
|
| 466 |
+
)
|
| 467 |
+
|
| 468 |
+
metadata_file = hf_hub_download(
|
| 469 |
+
repo_id=self.config.repo_id,
|
| 470 |
+
filename=self.config.metadata_file,
|
| 471 |
+
repo_type="dataset",
|
| 472 |
+
token=None # Forcer l'accès sans token pour les repos publics
|
| 473 |
+
)
|
| 474 |
+
except Exception as auth_error:
|
| 475 |
+
print(f" ⚠️ Erreur d'authentification: {auth_error}")
|
| 476 |
+
print(" 🔑 Essai avec token depuis les variables d'environnement...")
|
| 477 |
+
|
| 478 |
+
# Essayer avec le token d'environnement
|
| 479 |
+
import os
|
| 480 |
+
hf_token = os.getenv('HF_TOKEN') or os.getenv('HUGGINGFACE_HUB_TOKEN')
|
| 481 |
+
|
| 482 |
+
if hf_token:
|
| 483 |
+
print(" 🔑 Token trouvé, nouvel essai...")
|
| 484 |
+
embeddings_file = hf_hub_download(
|
| 485 |
+
repo_id=self.config.repo_id,
|
| 486 |
+
filename=self.config.embeddings_file,
|
| 487 |
+
repo_type="dataset",
|
| 488 |
+
token=hf_token
|
| 489 |
+
)
|
| 490 |
+
|
| 491 |
+
metadata_file = hf_hub_download(
|
| 492 |
+
repo_id=self.config.repo_id,
|
| 493 |
+
filename=self.config.metadata_file,
|
| 494 |
+
repo_type="dataset",
|
| 495 |
+
token=hf_token
|
| 496 |
+
)
|
| 497 |
+
else:
|
| 498 |
+
print(" ❌ Aucun token trouvé dans les variables d'environnement")
|
| 499 |
+
print(" 💡 Solutions possibles:")
|
| 500 |
+
print(" 1. Vérifiez que le repository est bien public")
|
| 501 |
+
print(" 2. Connectez-vous avec: huggingface-cli login")
|
| 502 |
+
print(" 3. Définissez HF_TOKEN dans les variables d'environnement")
|
| 503 |
+
raise auth_error
|
| 504 |
+
|
| 505 |
+
print(" 📥 Chargement des embeddings SafeTensors...")
|
| 506 |
+
tensors = load_file(embeddings_file)
|
| 507 |
+
embeddings_tensor = tensors["embeddings"]
|
| 508 |
+
embeddings_np = embeddings_tensor.numpy().astype(np.float32)
|
| 509 |
+
|
| 510 |
+
print(" 📋 Chargement des métadonnées...")
|
| 511 |
+
with open(metadata_file, 'r', encoding='utf-8') as f:
|
| 512 |
+
self.metadata = json.load(f)
|
| 513 |
+
|
| 514 |
+
# Créer l'index FAISS (optimisé pour Mac)
|
| 515 |
+
print(" 🔧 Création de l'index FAISS...")
|
| 516 |
+
dimension = embeddings_np.shape[1]
|
| 517 |
+
|
| 518 |
+
# Configuration d'index FAISS selon l'environnement
|
| 519 |
+
if self.is_zerogpu:
|
| 520 |
+
print(" 🚀 Index FAISS optimisé pour ZeroGPU (IndexHNSWFlat)")
|
| 521 |
+
# Index sophistiqué pour ZeroGPU avec GPU puissant
|
| 522 |
+
self.faiss_index = faiss.IndexHNSWFlat(dimension, 32)
|
| 523 |
+
self.faiss_index.hnsw.efConstruction = 200
|
| 524 |
+
self.faiss_index.hnsw.efSearch = 50
|
| 525 |
+
elif self.is_mps:
|
| 526 |
+
print(" 🍎 Index FAISS optimisé pour Mac (IndexFlatIP)")
|
| 527 |
+
# Index simple mais efficace sur Mac
|
| 528 |
+
self.faiss_index = faiss.IndexFlatIP(dimension) # Inner Product (plus stable sur Mac)
|
| 529 |
+
else:
|
| 530 |
+
print(" 🐧 Index FAISS HNSW pour Linux/Windows")
|
| 531 |
+
# Index plus sophistiqué pour autres plateformes
|
| 532 |
+
self.faiss_index = faiss.IndexHNSWFlat(dimension, 32)
|
| 533 |
+
self.faiss_index.hnsw.efConstruction = 200
|
| 534 |
+
self.faiss_index.hnsw.efSearch = 50
|
| 535 |
+
|
| 536 |
+
# Normaliser les embeddings pour IndexFlatIP (équivalent à cosine similarity)
|
| 537 |
+
if self.is_mps:
|
| 538 |
+
# Normalisation L2 pour que IndexFlatIP = cosine similarity
|
| 539 |
+
norms = np.linalg.norm(embeddings_np, axis=1, keepdims=True)
|
| 540 |
+
embeddings_np = embeddings_np / (norms + 1e-8) # Éviter division par 0
|
| 541 |
+
|
| 542 |
+
print(f" 📊 Ajout de {embeddings_np.shape[0]:,} vecteurs à l'index...")
|
| 543 |
+
# Ajouter les vecteurs à l'index
|
| 544 |
+
self.faiss_index.add(embeddings_np)
|
| 545 |
+
|
| 546 |
+
# Récupérer les mappings et métadonnées de contenu
|
| 547 |
+
self.ordered_ids = self.metadata.get('ordered_ids', [])
|
| 548 |
+
self.id_to_idx = self.metadata.get('id_to_idx', {})
|
| 549 |
+
self.content_metadata = self.metadata.get('content_metadata', {})
|
| 550 |
+
|
| 551 |
+
|
| 552 |
+
print(f"✅ Embeddings chargés: {embeddings_np.shape[0]:,} vecteurs de dimension {dimension}")
|
| 553 |
+
|
| 554 |
+
except Exception as e:
|
| 555 |
+
print(f"❌ Erreur lors du chargement des embeddings: {e}")
|
| 556 |
+
raise
|
| 557 |
+
|
| 558 |
+
def _load_embedding_model(self):
|
| 559 |
+
"""Charge le modèle d'embeddings pour les requêtes"""
|
| 560 |
+
print(f"🔄 Chargement du modèle d'embeddings {self.config.embedding_model}...")
|
| 561 |
+
|
| 562 |
+
try:
|
| 563 |
+
from sentence_transformers import SentenceTransformer
|
| 564 |
+
|
| 565 |
+
if self.use_flash_attention and self.is_cuda:
|
| 566 |
+
print(" - Configuration avec Flash Attention 2 activée (CUDA)")
|
| 567 |
+
try:
|
| 568 |
+
self.embedding_model = SentenceTransformer(
|
| 569 |
+
self.config.embedding_model,
|
| 570 |
+
model_kwargs={
|
| 571 |
+
"attn_implementation": "flash_attention_2",
|
| 572 |
+
"device_map": "auto"
|
| 573 |
+
},
|
| 574 |
+
tokenizer_kwargs={"padding_side": "left"}
|
| 575 |
+
)
|
| 576 |
+
except Exception as flash_error:
|
| 577 |
+
print(f" - Flash Attention échoué: {flash_error}")
|
| 578 |
+
print(" - Fallback vers configuration standard")
|
| 579 |
+
self.embedding_model = SentenceTransformer(self.config.embedding_model)
|
| 580 |
+
self.use_flash_attention = False
|
| 581 |
+
else:
|
| 582 |
+
print(" - Configuration standard (MPS/CPU ou Flash Attention désactivé)")
|
| 583 |
+
model_kwargs = {}
|
| 584 |
+
|
| 585 |
+
if self.is_mps:
|
| 586 |
+
model_kwargs = {"torch_dtype": torch.float32}
|
| 587 |
+
|
| 588 |
+
if model_kwargs:
|
| 589 |
+
self.embedding_model = SentenceTransformer(
|
| 590 |
+
self.config.embedding_model,
|
| 591 |
+
model_kwargs=model_kwargs,
|
| 592 |
+
tokenizer_kwargs={"padding_side": "left"}
|
| 593 |
+
)
|
| 594 |
+
else:
|
| 595 |
+
self.embedding_model = SentenceTransformer(self.config.embedding_model)
|
| 596 |
+
|
| 597 |
+
print(f"✅ Modèle d'embeddings {self.config.embedding_model} chargé avec succès")
|
| 598 |
+
|
| 599 |
+
except Exception as e:
|
| 600 |
+
print(f"❌ Erreur avec {self.config.embedding_model}: {e}")
|
| 601 |
+
print("🔄 Fallback vers le modèle multilingual MiniLM...")
|
| 602 |
+
self.embedding_model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')
|
| 603 |
+
self.use_flash_attention = False
|
| 604 |
+
|
| 605 |
+
def _load_reranker(self):
|
| 606 |
+
"""Charge le reranker Qwen3-Reranker-4B"""
|
| 607 |
+
if self.use_reranker:
|
| 608 |
+
try:
|
| 609 |
+
effective_flash_attention = self.use_flash_attention and not self.is_mps
|
| 610 |
+
self.reranker = Qwen3Reranker(use_flash_attention=effective_flash_attention)
|
| 611 |
+
except Exception as e:
|
| 612 |
+
print(f"❌ Erreur lors du chargement du reranker: {e}")
|
| 613 |
+
print("🔄 Désactivation du reranking")
|
| 614 |
+
self.use_reranker = False
|
| 615 |
+
self.reranker = None
|
| 616 |
+
else:
|
| 617 |
+
self.reranker = None
|
| 618 |
+
print("⚠️ Reranking désactivé par configuration")
|
| 619 |
+
|
| 620 |
+
def _load_generation_model(self):
|
| 621 |
+
"""Charge le modèle de génération Qwen3-4B-Instruct"""
|
| 622 |
+
print(f"🔄 Chargement du modèle de génération {self.generation_model_name}...")
|
| 623 |
+
|
| 624 |
+
try:
|
| 625 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 626 |
+
|
| 627 |
+
# Chargement du tokenizer
|
| 628 |
+
print(" - Chargement du tokenizer...")
|
| 629 |
+
self.generation_tokenizer = AutoTokenizer.from_pretrained(self.generation_model_name)
|
| 630 |
+
|
| 631 |
+
# Configuration du modèle selon la plateforme
|
| 632 |
+
model_kwargs = self._get_generation_model_config()
|
| 633 |
+
|
| 634 |
+
# Chargement du modèle
|
| 635 |
+
print(" - Chargement du modèle...")
|
| 636 |
+
self.generation_model = AutoModelForCausalLM.from_pretrained(
|
| 637 |
+
self.generation_model_name,
|
| 638 |
+
**model_kwargs
|
| 639 |
+
)
|
| 640 |
+
|
| 641 |
+
# Configuration du device
|
| 642 |
+
self._setup_generation_device()
|
| 643 |
+
|
| 644 |
+
print(f"✅ Modèle de génération chargé sur {self.generation_device}")
|
| 645 |
+
print(f" - Paramètres: {self._get_generation_parameter_count():.1f}B")
|
| 646 |
+
|
| 647 |
+
except Exception as e:
|
| 648 |
+
print(f"❌ Erreur lors du chargement du modèle de génération: {e}")
|
| 649 |
+
print("💡 La génération sera désactivée")
|
| 650 |
+
self.generation_model = None
|
| 651 |
+
self.generation_tokenizer = None
|
| 652 |
+
self.generation_device = None
|
| 653 |
+
|
| 654 |
+
def _get_generation_model_config(self) -> Dict:
|
| 655 |
+
"""Retourne la configuration du modèle de génération selon la plateforme"""
|
| 656 |
+
config = {}
|
| 657 |
+
|
| 658 |
+
if self.is_mps:
|
| 659 |
+
config["torch_dtype"] = torch.float32
|
| 660 |
+
config["device_map"] = None
|
| 661 |
+
elif self.is_cuda:
|
| 662 |
+
config["torch_dtype"] = torch.float16
|
| 663 |
+
if self.use_flash_attention:
|
| 664 |
+
try:
|
| 665 |
+
config["attn_implementation"] = "flash_attention_2"
|
| 666 |
+
print(" - Flash Attention 2 activée pour génération")
|
| 667 |
+
except Exception:
|
| 668 |
+
print(" - Flash Attention 2 non disponible pour génération")
|
| 669 |
+
config["device_map"] = "auto"
|
| 670 |
+
else:
|
| 671 |
+
config["torch_dtype"] = torch.float32
|
| 672 |
+
config["device_map"] = "cpu"
|
| 673 |
+
|
| 674 |
+
return config
|
| 675 |
+
|
| 676 |
+
def _setup_generation_device(self):
|
| 677 |
+
"""Configure le device pour le modèle de génération"""
|
| 678 |
+
if self.is_mps:
|
| 679 |
+
self.generation_device = torch.device("mps")
|
| 680 |
+
self.generation_model = self.generation_model.to(self.generation_device)
|
| 681 |
+
elif self.is_cuda:
|
| 682 |
+
if hasattr(self.generation_model, 'device'):
|
| 683 |
+
self.generation_device = next(self.generation_model.parameters()).device
|
| 684 |
+
else:
|
| 685 |
+
self.generation_device = torch.device("cuda")
|
| 686 |
+
self.generation_model = self.generation_model.to(self.generation_device)
|
| 687 |
+
else:
|
| 688 |
+
self.generation_device = torch.device("cpu")
|
| 689 |
+
self.generation_model = self.generation_model.to(self.generation_device)
|
| 690 |
+
|
| 691 |
+
def _get_generation_parameter_count(self) -> float:
|
| 692 |
+
"""Retourne le nombre de paramètres du modèle de génération en milliards"""
|
| 693 |
+
if self.generation_model is None:
|
| 694 |
+
return 0.0
|
| 695 |
+
try:
|
| 696 |
+
return sum(p.numel() for p in self.generation_model.parameters()) / 1e9
|
| 697 |
+
except:
|
| 698 |
+
return 0.0
|
| 699 |
+
|
| 700 |
+
def search_documents(self, query: str, final_k: int = None, use_reranking: bool = None) -> List[Dict]:
|
| 701 |
+
"""
|
| 702 |
+
Recherche avancée avec reranking en deux étapes
|
| 703 |
+
"""
|
| 704 |
+
k = final_k if final_k is not None else self.final_k
|
| 705 |
+
initial_k = max(self.initial_k, k * 3)
|
| 706 |
+
should_rerank = use_reranking if use_reranking is not None else self.use_reranker
|
| 707 |
+
|
| 708 |
+
print(f"🔍 Recherche en deux étapes: {initial_k} candidats → reranking → {k} finaux")
|
| 709 |
+
|
| 710 |
+
# Étape 1: Recherche par embedding avec FAISS
|
| 711 |
+
if hasattr(self.embedding_model, 'prompts') and 'query' in self.embedding_model.prompts:
|
| 712 |
+
query_embedding = self.embedding_model.encode([query], prompt_name="query")[0]
|
| 713 |
+
else:
|
| 714 |
+
query_embedding = self.embedding_model.encode([query])[0]
|
| 715 |
+
|
| 716 |
+
# Recherche dans l'index FAISS
|
| 717 |
+
query_vector = query_embedding.reshape(1, -1).astype('float32')
|
| 718 |
+
|
| 719 |
+
# Normaliser la requête sur Mac pour IndexFlatIP (consistency avec les embeddings)
|
| 720 |
+
if self.is_mps:
|
| 721 |
+
norm = np.linalg.norm(query_vector)
|
| 722 |
+
if norm > 0:
|
| 723 |
+
query_vector = query_vector / norm
|
| 724 |
+
|
| 725 |
+
distances, indices = self.faiss_index.search(query_vector, initial_k)
|
| 726 |
+
|
| 727 |
+
if len(indices[0]) == 0:
|
| 728 |
+
print("❌ Aucun document trouvé")
|
| 729 |
+
return []
|
| 730 |
+
|
| 731 |
+
print(f"📋 {len(indices[0])} candidats récupérés")
|
| 732 |
+
|
| 733 |
+
# Conversion en format intermédiaire
|
| 734 |
+
initial_results = []
|
| 735 |
+
for i, (distance, idx) in enumerate(zip(distances[0], indices[0])):
|
| 736 |
+
if idx < len(self.ordered_ids):
|
| 737 |
+
doc_id = self.ordered_ids[idx]
|
| 738 |
+
doc_metadata = self.content_metadata.get(doc_id, {})
|
| 739 |
+
|
| 740 |
+
# Ajustement des scores selon le type d'index
|
| 741 |
+
if self.is_mps:
|
| 742 |
+
# Sur Mac avec IndexFlatIP : distance = inner product (plus haut = plus similaire)
|
| 743 |
+
embedding_score = float(distance) # Inner product normalisé = cosine similarity
|
| 744 |
+
embedding_distance = 1.0 - embedding_score # Conversion en distance pour compatibilité
|
| 745 |
+
else:
|
| 746 |
+
# Sur autres plateformes avec IndexHNSWFlat : distance euclidienne
|
| 747 |
+
embedding_distance = float(distance)
|
| 748 |
+
embedding_score = 1 - embedding_distance
|
| 749 |
+
|
| 750 |
+
doc = {
|
| 751 |
+
'content': doc_metadata.get('chunk_content', 'Contenu non disponible'),
|
| 752 |
+
'metadata': doc_metadata,
|
| 753 |
+
'embedding_distance': embedding_distance,
|
| 754 |
+
'embedding_score': embedding_score,
|
| 755 |
+
'source': doc_metadata.get('source_file', 'Inconnu'),
|
| 756 |
+
'title': doc_metadata.get('title', 'Sans titre'),
|
| 757 |
+
'heading': doc_metadata.get('heading', ''),
|
| 758 |
+
'initial_rank': i + 1
|
| 759 |
+
}
|
| 760 |
+
initial_results.append(doc)
|
| 761 |
+
|
| 762 |
+
# Étape 2: Reranking si disponible
|
| 763 |
+
if should_rerank and self.reranker and self.reranker.model is not None:
|
| 764 |
+
print("🎯 Application du reranking Qwen3...")
|
| 765 |
+
|
| 766 |
+
documents = [doc['content'] for doc in initial_results]
|
| 767 |
+
|
| 768 |
+
|
| 769 |
+
rerank_scores = self.reranker.rerank(query, documents)
|
| 770 |
+
|
| 771 |
+
# Ajout des scores de reranking
|
| 772 |
+
for doc, rerank_score in zip(initial_results, rerank_scores):
|
| 773 |
+
doc['rerank_score'] = float(rerank_score)
|
| 774 |
+
|
| 775 |
+
# Tri par score de reranking
|
| 776 |
+
initial_results.sort(key=lambda x: x['rerank_score'], reverse=True)
|
| 777 |
+
|
| 778 |
+
# Mise à jour des positions finales
|
| 779 |
+
for i, doc in enumerate(initial_results):
|
| 780 |
+
doc['final_rank'] = i + 1
|
| 781 |
+
|
| 782 |
+
top_scores = [f"{doc['rerank_score']:.3f}" for doc in initial_results[:5]]
|
| 783 |
+
print(f"✅ Reranking appliqué, top 5 scores: {top_scores}")
|
| 784 |
+
else:
|
| 785 |
+
print("⚠️ Reranking désactivé, utilisation des scores d'embedding uniquement")
|
| 786 |
+
for doc in initial_results:
|
| 787 |
+
doc['rerank_score'] = doc['embedding_score']
|
| 788 |
+
doc['final_rank'] = doc['initial_rank']
|
| 789 |
+
|
| 790 |
+
# Retour des top-k résultats finaux
|
| 791 |
+
final_results = initial_results[:k]
|
| 792 |
+
print(f"📊 {len(final_results)} documents finaux sélectionnés")
|
| 793 |
+
|
| 794 |
+
return final_results
|
| 795 |
+
|
| 796 |
+
def generate_response_stream(self, query: str, context: str, history: List = None):
|
| 797 |
+
"""
|
| 798 |
+
Génère une réponse streamée basée sur le contexte et l'historique
|
| 799 |
+
"""
|
| 800 |
+
if self.generation_model is None or self.generation_tokenizer is None:
|
| 801 |
+
yield "❌ Modèle de génération non disponible"
|
| 802 |
+
return
|
| 803 |
+
|
| 804 |
+
# Construction du prompt système
|
| 805 |
+
system_prompt = """Tu es un assistant expert qui répond aux questions en te basant uniquement sur les documents fournis dans le contexte.
|
| 806 |
+
|
| 807 |
+
Instructions importantes:
|
| 808 |
+
- Réponds en français de manière claire et précise
|
| 809 |
+
- Base-toi uniquement sur les informations du contexte fourni
|
| 810 |
+
- Si l'information n'est pas dans le contexte, dis-le clairement
|
| 811 |
+
- Utilise un ton professionnel adapté au domaine
|
| 812 |
+
- Structure ta réponse avec des paragraphes clairs"""
|
| 813 |
+
|
| 814 |
+
# Construire le prompt complet
|
| 815 |
+
messages = [{"role": "system", "content": system_prompt}]
|
| 816 |
+
|
| 817 |
+
# Ajouter l'historique si fourni
|
| 818 |
+
if history:
|
| 819 |
+
for msg in history:
|
| 820 |
+
if hasattr(msg, 'role') and hasattr(msg, 'content'):
|
| 821 |
+
messages.append({"role": msg.role, "content": msg.content})
|
| 822 |
+
|
| 823 |
+
# Ajouter le contexte et la question
|
| 824 |
+
user_message = f"Contexte:\n{context}\n\nQuestion: {query}"
|
| 825 |
+
messages.append({"role": "user", "content": user_message})
|
| 826 |
+
|
| 827 |
+
try:
|
| 828 |
+
# Tokenisation
|
| 829 |
+
inputs = self.generation_tokenizer.apply_chat_template(
|
| 830 |
+
messages,
|
| 831 |
+
tokenize=True,
|
| 832 |
+
add_generation_prompt=True,
|
| 833 |
+
return_tensors="pt"
|
| 834 |
+
).to(self.device)
|
| 835 |
+
|
| 836 |
+
# Génération streamée
|
| 837 |
+
from transformers import TextIteratorStreamer
|
| 838 |
+
import threading
|
| 839 |
+
|
| 840 |
+
streamer = TextIteratorStreamer(
|
| 841 |
+
self.generation_tokenizer,
|
| 842 |
+
timeout=10.0,
|
| 843 |
+
skip_prompt=True,
|
| 844 |
+
skip_special_tokens=True
|
| 845 |
+
)
|
| 846 |
+
|
| 847 |
+
generation_kwargs = {
|
| 848 |
+
"input_ids": inputs,
|
| 849 |
+
"streamer": streamer,
|
| 850 |
+
"max_new_tokens": 1024,
|
| 851 |
+
"temperature": 0.7,
|
| 852 |
+
"do_sample": True,
|
| 853 |
+
"pad_token_id": self.generation_tokenizer.eos_token_id,
|
| 854 |
+
"eos_token_id": self.generation_tokenizer.eos_token_id,
|
| 855 |
+
}
|
| 856 |
+
|
| 857 |
+
# Lancer la génération dans un thread séparé
|
| 858 |
+
thread = threading.Thread(target=self.generation_model.generate, kwargs=generation_kwargs)
|
| 859 |
+
thread.start()
|
| 860 |
+
|
| 861 |
+
# Streamer les tokens
|
| 862 |
+
for new_token in streamer:
|
| 863 |
+
yield new_token
|
| 864 |
+
|
| 865 |
+
thread.join()
|
| 866 |
+
|
| 867 |
+
except Exception as e:
|
| 868 |
+
yield f"❌ Erreur lors de la génération: {str(e)}"
|
| 869 |
+
|
| 870 |
+
def generate_response(self, query: str, context: str, history: List = None) -> str:
|
| 871 |
+
"""
|
| 872 |
+
Génère une réponse basée sur le contexte et l'historique
|
| 873 |
+
"""
|
| 874 |
+
if self.generation_model is None or self.generation_tokenizer is None:
|
| 875 |
+
return "❌ Modèle de génération non disponible"
|
| 876 |
+
|
| 877 |
+
# Construction du prompt système
|
| 878 |
+
system_prompt = """Tu es un assistant expert qui répond aux questions en te basant uniquement sur les documents fournis dans le contexte.
|
| 879 |
+
|
| 880 |
+
Instructions importantes:
|
| 881 |
+
- Réponds en français de manière claire et précise
|
| 882 |
+
- Base-toi uniquement sur les informations du contexte fourni
|
| 883 |
+
- Si l'information n'est pas dans le contexte, dis-le clairement
|
| 884 |
+
- Utilise un ton professionnel adapté au domaine
|
| 885 |
+
- Structure ta réponse avec des paragraphes clairs"""
|
| 886 |
+
|
| 887 |
+
# Construire le prompt complet
|
| 888 |
+
messages = [{"role": "system", "content": system_prompt}]
|
| 889 |
+
|
| 890 |
+
# Ajouter l'historique si fourni
|
| 891 |
+
if history:
|
| 892 |
+
for msg in history:
|
| 893 |
+
if hasattr(msg, 'role') and hasattr(msg, 'content'):
|
| 894 |
+
if msg.role in ["user", "assistant"] and not getattr(msg, 'metadata', None):
|
| 895 |
+
messages.append({"role": msg.role, "content": msg.content})
|
| 896 |
+
|
| 897 |
+
# Ajouter la question courante avec le contexte
|
| 898 |
+
user_prompt = f"""Contexte documentaire:
|
| 899 |
+
{context}
|
| 900 |
+
|
| 901 |
+
Question: {query}
|
| 902 |
+
|
| 903 |
+
Réponds à cette question en te basant sur le contexte fourni."""
|
| 904 |
+
|
| 905 |
+
messages.append({"role": "user", "content": user_prompt})
|
| 906 |
+
|
| 907 |
+
# Formatage pour le modèle
|
| 908 |
+
try:
|
| 909 |
+
# Appliquer le template de chat du modèle
|
| 910 |
+
formatted_prompt = self.generation_tokenizer.apply_chat_template(
|
| 911 |
+
messages,
|
| 912 |
+
tokenize=False,
|
| 913 |
+
add_generation_prompt=True
|
| 914 |
+
)
|
| 915 |
+
|
| 916 |
+
# Tokenisation
|
| 917 |
+
inputs = self.generation_tokenizer(
|
| 918 |
+
formatted_prompt,
|
| 919 |
+
return_tensors="pt",
|
| 920 |
+
truncation=True,
|
| 921 |
+
max_length=4096
|
| 922 |
+
)
|
| 923 |
+
|
| 924 |
+
# Déplacement vers le device
|
| 925 |
+
inputs = {k: v.to(self.generation_device) for k, v in inputs.items()}
|
| 926 |
+
|
| 927 |
+
# Génération
|
| 928 |
+
with torch.no_grad():
|
| 929 |
+
outputs = self.generation_model.generate(
|
| 930 |
+
**inputs,
|
| 931 |
+
max_new_tokens=1024,
|
| 932 |
+
temperature=0.7,
|
| 933 |
+
do_sample=True,
|
| 934 |
+
pad_token_id=self.generation_tokenizer.eos_token_id,
|
| 935 |
+
eos_token_id=self.generation_tokenizer.eos_token_id,
|
| 936 |
+
)
|
| 937 |
+
|
| 938 |
+
# Décodage de la réponse
|
| 939 |
+
full_response = self.generation_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 940 |
+
|
| 941 |
+
# Extraire seulement la nouvelle génération
|
| 942 |
+
response = full_response[len(formatted_prompt):].strip()
|
| 943 |
+
|
| 944 |
+
return response
|
| 945 |
+
|
| 946 |
+
except Exception as e:
|
| 947 |
+
print(f"❌ Erreur lors de la génération: {e}")
|
| 948 |
+
return f"❌ Erreur lors de la génération de la réponse: {str(e)}"
|
| 949 |
+
|
| 950 |
+
def stream_response_with_tools(self, query: str, history, top_k: int = None, use_reranking: bool = None):
|
| 951 |
+
"""
|
| 952 |
+
Génère une réponse streamée avec affichage visuel des tools et reranking Qwen3
|
| 953 |
+
"""
|
| 954 |
+
# 1. S'assurer que l'historique est une liste
|
| 955 |
+
if not history:
|
| 956 |
+
history = []
|
| 957 |
+
|
| 958 |
+
# 2. Ajouter le message utilisateur seulement s'il n'est pas déjà présent
|
| 959 |
+
if not history or history[-1].role != "user" or history[-1].content != query:
|
| 960 |
+
history.append(ChatMessage(role="user", content=query))
|
| 961 |
+
yield history
|
| 962 |
+
time.sleep(0.1)
|
| 963 |
+
|
| 964 |
+
# 3. Recherche des documents avec tool visuel
|
| 965 |
+
should_rerank = use_reranking if use_reranking is not None else self.use_reranker
|
| 966 |
+
search_method = "avec reranking Qwen3" if should_rerank else "par embedding seulement"
|
| 967 |
+
|
| 968 |
+
history.append(ChatMessage(
|
| 969 |
+
role="assistant",
|
| 970 |
+
content=f"Je recherche les documents les plus pertinents dans la base de données ({search_method})...",
|
| 971 |
+
metadata={"title": "🔍 Recherche sémantique avancée"}
|
| 972 |
+
))
|
| 973 |
+
yield history
|
| 974 |
+
|
| 975 |
+
# Recherche des documents pertinents
|
| 976 |
+
relevant_docs = self.search_documents(query, top_k, use_reranking)
|
| 977 |
+
|
| 978 |
+
time.sleep(0.2)
|
| 979 |
+
|
| 980 |
+
if not relevant_docs:
|
| 981 |
+
history.append(ChatMessage(
|
| 982 |
+
role="assistant",
|
| 983 |
+
content="Aucun document pertinent trouvé dans la base de données."
|
| 984 |
+
))
|
| 985 |
+
yield history
|
| 986 |
+
return
|
| 987 |
+
|
| 988 |
+
# 4. Affichage des documents trouvés avec scores détaillés
|
| 989 |
+
docs_summary = f"Trouvé {len(relevant_docs)} documents pertinents"
|
| 990 |
+
if should_rerank:
|
| 991 |
+
docs_summary += f"\n\n📊 **Reranking Qwen3 appliqué:**"
|
| 992 |
+
for i, doc in enumerate(relevant_docs):
|
| 993 |
+
embedding_score = doc.get('embedding_score', 0)
|
| 994 |
+
rerank_score = doc.get('rerank_score', 0)
|
| 995 |
+
rank_change = doc.get('initial_rank', i+1) - doc.get('final_rank', i+1)
|
| 996 |
+
rank_indicator = f" (#{doc.get('initial_rank', i+1)}→#{doc.get('final_rank', i+1)})" if rank_change != 0 else ""
|
| 997 |
+
docs_summary += f"\n• **{doc['title']}**{rank_indicator}"
|
| 998 |
+
docs_summary += f"\n └ Embedding: {embedding_score:.3f} | Reranking: {rerank_score:.3f}"
|
| 999 |
+
else:
|
| 1000 |
+
for i, doc in enumerate(relevant_docs):
|
| 1001 |
+
embedding_score = doc.get('embedding_score', doc.get('distance', 0))
|
| 1002 |
+
docs_summary += f"\n• **{doc['title']}** - Score: {embedding_score:.3f}"
|
| 1003 |
+
|
| 1004 |
+
history.append(ChatMessage(
|
| 1005 |
+
role="assistant",
|
| 1006 |
+
content=docs_summary,
|
| 1007 |
+
metadata={"title": f"📚 Documents sélectionnés ({len(relevant_docs)} total)"}
|
| 1008 |
+
))
|
| 1009 |
+
yield history
|
| 1010 |
+
|
| 1011 |
+
time.sleep(0.2)
|
| 1012 |
+
|
| 1013 |
+
# 5. Construction du contexte
|
| 1014 |
+
context_parts = []
|
| 1015 |
+
sources_with_scores = []
|
| 1016 |
+
|
| 1017 |
+
for i, doc in enumerate(relevant_docs):
|
| 1018 |
+
context_parts.append(f"[Document {i+1}] {doc['title']} - {doc['heading']}\n{doc['content']}")
|
| 1019 |
+
sources_with_scores.append({
|
| 1020 |
+
'title': doc['title'],
|
| 1021 |
+
'source': doc['source'],
|
| 1022 |
+
'embedding_score': doc.get('embedding_score', 1 - doc.get('distance', 0)),
|
| 1023 |
+
'rerank_score': doc.get('rerank_score'),
|
| 1024 |
+
'final_rank': doc.get('final_rank', i+1)
|
| 1025 |
+
})
|
| 1026 |
+
|
| 1027 |
+
context = "\n\n".join(context_parts)
|
| 1028 |
+
|
| 1029 |
+
# 6. Génération de la réponse avec Qwen3-4B
|
| 1030 |
+
history.append(ChatMessage(
|
| 1031 |
+
role="assistant",
|
| 1032 |
+
content="Génération de la réponse basée sur les documents sélectionnés...",
|
| 1033 |
+
metadata={"title": "🤖 Génération avec Qwen3-4B"}
|
| 1034 |
+
))
|
| 1035 |
+
yield history
|
| 1036 |
+
|
| 1037 |
+
time.sleep(0.2)
|
| 1038 |
+
|
| 1039 |
+
# Génération streamée de la réponse
|
| 1040 |
+
history.append(ChatMessage(
|
| 1041 |
+
role="assistant",
|
| 1042 |
+
content="", # Commencer avec un contenu vide
|
| 1043 |
+
metadata={"title": "🤖 Réponse générée"}
|
| 1044 |
+
))
|
| 1045 |
+
|
| 1046 |
+
# Streamer la réponse token par token
|
| 1047 |
+
current_response = ""
|
| 1048 |
+
for token in self.generate_response_stream(query, context, history[:-1]): # Exclure le dernier message vide
|
| 1049 |
+
current_response += token
|
| 1050 |
+
# Mettre à jour le dernier message avec la réponse en cours
|
| 1051 |
+
history[-1] = ChatMessage(
|
| 1052 |
+
role="assistant",
|
| 1053 |
+
content=current_response,
|
| 1054 |
+
metadata={"title": "🤖 Réponse générée"}
|
| 1055 |
+
)
|
| 1056 |
+
yield history
|
| 1057 |
+
time.sleep(0.01) # Petit délai pour un streaming fluide
|
| 1058 |
+
|
| 1059 |
+
time.sleep(0.2)
|
| 1060 |
+
|
| 1061 |
+
# 7. Ajout des sources consultées avec scores détaillés
|
| 1062 |
+
sources_text = []
|
| 1063 |
+
for i, source_info in enumerate(sources_with_scores):
|
| 1064 |
+
embedding_score = source_info['embedding_score']
|
| 1065 |
+
rerank_score = source_info.get('rerank_score')
|
| 1066 |
+
source_file = source_info['source']
|
| 1067 |
+
|
| 1068 |
+
if rerank_score is not None:
|
| 1069 |
+
score_display = f"Embedding: {embedding_score:.3f} | **Reranking: {rerank_score:.3f}**"
|
| 1070 |
+
else:
|
| 1071 |
+
score_display = f"Score: {embedding_score:.3f}"
|
| 1072 |
+
|
| 1073 |
+
sources_text.append(f"• **[{i+1}]** {source_info['title']} ({source_file})\n └ {score_display}")
|
| 1074 |
+
|
| 1075 |
+
sources_display = "\n".join(sources_text)
|
| 1076 |
+
|
| 1077 |
+
# Titre adaptatif selon la méthode utilisée
|
| 1078 |
+
sources_title = f"📚 Sources avec reranking Qwen3 ({len(relevant_docs)} documents)" if should_rerank else f"📚 Sources par embedding ({len(relevant_docs)} documents)"
|
| 1079 |
+
|
| 1080 |
+
history.append(ChatMessage(
|
| 1081 |
+
role="assistant",
|
| 1082 |
+
content=sources_display,
|
| 1083 |
+
metadata={"title": sources_title}
|
| 1084 |
+
))
|
| 1085 |
+
yield history
|
| 1086 |
+
|
| 1087 |
+
|
| 1088 |
+
def _create_rag_system():
|
| 1089 |
+
"""Créé et configure le système RAG avec paramètres optimaux"""
|
| 1090 |
+
|
| 1091 |
+
# Détection automatique d'environnement
|
| 1092 |
+
is_zerogpu = ZEROGPU_AVAILABLE and os.getenv("SPACE_ID") is not None
|
| 1093 |
+
is_mac = torch.backends.mps.is_available() and not is_zerogpu
|
| 1094 |
+
is_cuda = torch.cuda.is_available()
|
| 1095 |
+
|
| 1096 |
+
if is_zerogpu:
|
| 1097 |
+
print("🚀 ZeroGPU détecté - optimisations cloud appliquées")
|
| 1098 |
+
elif is_mac:
|
| 1099 |
+
print("🍎 Mac avec MPS détecté - optimisations automatiques appliquées")
|
| 1100 |
+
elif is_cuda:
|
| 1101 |
+
print("🐧 CUDA détecté - optimisations GPU appliquées")
|
| 1102 |
+
else:
|
| 1103 |
+
print("💻 CPU détecté - optimisations processeur appliquées")
|
| 1104 |
+
|
| 1105 |
+
# Paramètres par défaut optimisés selon l'environnement
|
| 1106 |
+
if is_zerogpu:
|
| 1107 |
+
default_config = {
|
| 1108 |
+
'use_flash_attention': True, # ZeroGPU supporte Flash Attention
|
| 1109 |
+
'use_reranker': True, # GPU puissant, reranking activé
|
| 1110 |
+
'initial_k': 30, # Plus de candidats avec GPU puissant
|
| 1111 |
+
'final_k': 5 # Plus de documents finaux
|
| 1112 |
+
}
|
| 1113 |
+
elif is_mac:
|
| 1114 |
+
default_config = {
|
| 1115 |
+
'use_flash_attention': False, # MPS ne supporte pas Flash Attention
|
| 1116 |
+
'use_reranker': True, # Reranking OK sur Mac
|
| 1117 |
+
'initial_k': 20, # Valeurs modérées
|
| 1118 |
+
'final_k': 3
|
| 1119 |
+
}
|
| 1120 |
+
else:
|
| 1121 |
+
default_config = {
|
| 1122 |
+
'use_flash_attention': is_cuda, # Flash Attention seulement sur CUDA
|
| 1123 |
+
'use_reranker': True, # Reranking par défaut
|
| 1124 |
+
'initial_k': 20, # Candidats pour la première étape
|
| 1125 |
+
'final_k': 3 # Documents finaux par défaut
|
| 1126 |
+
}
|
| 1127 |
+
|
| 1128 |
+
print("🚀 Initialisation du chatbot RAG générique...")
|
| 1129 |
+
return GenericRAGChatbot(**default_config)
|
| 1130 |
+
|
| 1131 |
+
|
| 1132 |
+
def _clear_message():
|
| 1133 |
+
"""Fonction utilitaire interne pour effacer le message d'entrée."""
|
| 1134 |
+
return ""
|
| 1135 |
+
|
| 1136 |
+
def _clear_chat():
|
| 1137 |
+
"""Fonction utilitaire interne pour effacer l'historique de chat."""
|
| 1138 |
+
return []
|
| 1139 |
+
|
| 1140 |
+
def _ensure_chatmessages(history):
|
| 1141 |
+
"""Convertit une liste en objets ChatMessage si besoin."""
|
| 1142 |
+
result = []
|
| 1143 |
+
for m in history or []:
|
| 1144 |
+
if isinstance(m, ChatMessage):
|
| 1145 |
+
result.append(m)
|
| 1146 |
+
elif isinstance(m, dict):
|
| 1147 |
+
result.append(ChatMessage(
|
| 1148 |
+
role=m.get("role", ""),
|
| 1149 |
+
content=m.get("content", ""),
|
| 1150 |
+
metadata=m.get("metadata", None)
|
| 1151 |
+
))
|
| 1152 |
+
elif isinstance(m, (list, tuple)) and len(m) >= 2:
|
| 1153 |
+
result.append(ChatMessage(role=m[0], content=m[1]))
|
| 1154 |
+
return result
|
| 1155 |
+
|
| 1156 |
+
|
| 1157 |
+
@spaces.GPU(duration=180) # ZeroGPU: alloue GPU pour toute la pipeline
|
| 1158 |
+
def chat_with_generic_rag(message, history, top_k, use_reranking):
|
| 1159 |
+
"""
|
| 1160 |
+
Interface entre Gradio et le système RAG générique avec contrôles avancés.
|
| 1161 |
+
|
| 1162 |
+
Cette fonction gère l'interface de chat interactive avec streaming en temps réel
|
| 1163 |
+
et affichage des étapes de traitement (recherche, reranking, génération).
|
| 1164 |
+
|
| 1165 |
+
Args:
|
| 1166 |
+
message (str): Le message ou question de l'utilisateur à traiter
|
| 1167 |
+
history (list): L'historique de la conversation sous forme de liste de messages
|
| 1168 |
+
top_k (int): Nombre de documents finaux à utiliser pour la génération de réponse
|
| 1169 |
+
use_reranking (bool): Activation du reranking Qwen3 pour améliorer la sélection
|
| 1170 |
+
|
| 1171 |
+
Yields:
|
| 1172 |
+
list: Historique mis à jour avec les nouveaux messages et étapes de traitement
|
| 1173 |
+
"""
|
| 1174 |
+
history = _ensure_chatmessages(history)
|
| 1175 |
+
response_generator = rag_system.stream_response_with_tools(message, history, top_k, use_reranking)
|
| 1176 |
+
for updated_history in response_generator:
|
| 1177 |
+
yield updated_history
|
| 1178 |
+
|
| 1179 |
+
|
| 1180 |
+
def ask_rag_question(question: str = "Qu'est-ce que Swift MLX?", num_documents: int = 3, use_reranking: bool = True) -> str:
|
| 1181 |
+
"""
|
| 1182 |
+
Pose une question au système RAG LocalRAG et retourne la réponse avec les documents sources.
|
| 1183 |
+
|
| 1184 |
+
Cette fonction utilise un système de recherche sémantique avancé avec des modèles Qwen3
|
| 1185 |
+
pour interroger une base de connaissances et générer des réponses contextualisées.
|
| 1186 |
+
|
| 1187 |
+
Args:
|
| 1188 |
+
question (str): La question à poser au système RAG en langage naturel
|
| 1189 |
+
num_documents (int): Nombre de documents à utiliser pour générer la réponse (entre 1 et 10)
|
| 1190 |
+
use_reranking (bool): Utiliser le reranking Qwen3-Reranker-4B pour améliorer la sélection des documents
|
| 1191 |
+
|
| 1192 |
+
Returns:
|
| 1193 |
+
str: Réponse générée incluant la réponse contextuelle et les sources avec leurs scores de pertinence
|
| 1194 |
+
"""
|
| 1195 |
+
global rag_system
|
| 1196 |
+
|
| 1197 |
+
try:
|
| 1198 |
+
# Validation des paramètres
|
| 1199 |
+
num_documents = max(1, min(10, int(num_documents)))
|
| 1200 |
+
|
| 1201 |
+
print(f"🔍 Question MCP: {question}")
|
| 1202 |
+
print(f"📊 Paramètres: {num_documents} documents, reranking: {use_reranking}")
|
| 1203 |
+
|
| 1204 |
+
# Recherche des documents pertinents
|
| 1205 |
+
relevant_docs = rag_system.search_documents(question, num_documents, use_reranking)
|
| 1206 |
+
|
| 1207 |
+
if not relevant_docs:
|
| 1208 |
+
return "❌ Aucun document pertinent trouvé dans la base de données pour répondre à cette question."
|
| 1209 |
+
|
| 1210 |
+
# Construction du contexte pour la génération
|
| 1211 |
+
context_parts = []
|
| 1212 |
+
for i, doc in enumerate(relevant_docs):
|
| 1213 |
+
context_parts.append(f"[Document {i+1}] {doc['title']} - {doc['heading']}\n{doc['content']}")
|
| 1214 |
+
|
| 1215 |
+
context = "\n\n".join(context_parts)
|
| 1216 |
+
|
| 1217 |
+
# Génération de la réponse
|
| 1218 |
+
response = rag_system.generate_response(question, context, None)
|
| 1219 |
+
|
| 1220 |
+
# Formatage de la réponse avec les sources
|
| 1221 |
+
sources_info = []
|
| 1222 |
+
search_method = "avec reranking Qwen3" if use_reranking else "par embedding seulement"
|
| 1223 |
+
|
| 1224 |
+
sources_info.append(f"\n\n📚 **Documents sources utilisés ({search_method}):**\n")
|
| 1225 |
+
|
| 1226 |
+
for i, doc in enumerate(relevant_docs):
|
| 1227 |
+
embedding_score = doc.get('embedding_score', 0)
|
| 1228 |
+
rerank_score = doc.get('rerank_score')
|
| 1229 |
+
initial_rank = doc.get('initial_rank', i+1)
|
| 1230 |
+
final_rank = doc.get('final_rank', i+1)
|
| 1231 |
+
|
| 1232 |
+
# Formatage des scores
|
| 1233 |
+
if rerank_score is not None and use_reranking:
|
| 1234 |
+
score_display = f"Embedding: {embedding_score:.3f} | **Reranking: {rerank_score:.3f}**"
|
| 1235 |
+
if initial_rank != final_rank:
|
| 1236 |
+
rank_change = f" (#{initial_rank}→#{final_rank})"
|
| 1237 |
+
else:
|
| 1238 |
+
rank_change = ""
|
| 1239 |
+
else:
|
| 1240 |
+
score_display = f"Score: {embedding_score:.3f}"
|
| 1241 |
+
rank_change = ""
|
| 1242 |
+
|
| 1243 |
+
sources_info.append(f"• **[{i+1}]** {doc['title']}{rank_change}")
|
| 1244 |
+
sources_info.append(f" └ {score_display}")
|
| 1245 |
+
sources_info.append(f" └ Source: {doc['source']}")
|
| 1246 |
+
|
| 1247 |
+
# Assemblage de la réponse finale
|
| 1248 |
+
final_response = response + "\n".join(sources_info)
|
| 1249 |
+
|
| 1250 |
+
print(f"✅ Réponse MCP générée ({len(relevant_docs)} documents utilisés)")
|
| 1251 |
+
return final_response
|
| 1252 |
+
|
| 1253 |
+
except Exception as e:
|
| 1254 |
+
error_msg = f"❌ Erreur lors du traitement de la question: {str(e)}"
|
| 1255 |
+
print(error_msg)
|
| 1256 |
+
return error_msg
|
| 1257 |
+
|
| 1258 |
+
|
| 1259 |
+
def create_gradio_interface():
|
| 1260 |
+
"""Créé l'interface Gradio pour utilisation externe (Spaces)"""
|
| 1261 |
+
# Initialisation du système RAG
|
| 1262 |
+
global rag_system
|
| 1263 |
+
try:
|
| 1264 |
+
rag_system = _create_rag_system()
|
| 1265 |
+
except Exception as e:
|
| 1266 |
+
raise RuntimeError(f"Erreur d'initialisation RAG: {e}")
|
| 1267 |
+
|
| 1268 |
+
# Configuration de l'interface Gradio avec thème Glass
|
| 1269 |
+
with gr.Blocks(
|
| 1270 |
+
title="🤖 LocalRAG Chat Générique",
|
| 1271 |
+
theme=gr.themes.Glass(),
|
| 1272 |
+
) as demo:
|
| 1273 |
+
|
| 1274 |
+
# En-tête simplifié avec composants Gradio natifs
|
| 1275 |
+
with gr.Row():
|
| 1276 |
+
with gr.Column():
|
| 1277 |
+
gr.Markdown("# 🤖 Assistant RAG Générique LocalRAG")
|
| 1278 |
+
gr.Markdown(f"📦 Repository: `{rag_system.config.repo_id}` | 📊 Vecteurs: **{rag_system.config.total_vectors:,}**")
|
| 1279 |
+
|
| 1280 |
+
with gr.Row():
|
| 1281 |
+
with gr.Column(scale=4):
|
| 1282 |
+
chatbot = gr.Chatbot(
|
| 1283 |
+
label="💬 Conversation avec l'assistant",
|
| 1284 |
+
show_label=True,
|
| 1285 |
+
height=600,
|
| 1286 |
+
type="messages"
|
| 1287 |
+
)
|
| 1288 |
+
|
| 1289 |
+
msg = gr.Textbox(
|
| 1290 |
+
label="Votre question",
|
| 1291 |
+
placeholder="Posez votre question ici...",
|
| 1292 |
+
lines=1,
|
| 1293 |
+
max_lines=3
|
| 1294 |
+
)
|
| 1295 |
+
|
| 1296 |
+
with gr.Row():
|
| 1297 |
+
send_btn = gr.Button("Envoyer", variant="primary")
|
| 1298 |
+
clear_btn = gr.Button("Effacer", variant="secondary")
|
| 1299 |
+
|
| 1300 |
+
with gr.Column(scale=1):
|
| 1301 |
+
gr.Markdown("### ⚙️ Paramètres")
|
| 1302 |
+
top_k_slider = gr.Slider(
|
| 1303 |
+
minimum=1,
|
| 1304 |
+
maximum=20,
|
| 1305 |
+
value=5,
|
| 1306 |
+
step=1,
|
| 1307 |
+
label="Nombre de documents (top-k)",
|
| 1308 |
+
info="Plus élevé = plus de contexte"
|
| 1309 |
+
)
|
| 1310 |
+
|
| 1311 |
+
reranking_checkbox = gr.Checkbox(
|
| 1312 |
+
label="Activer reranking Qwen3",
|
| 1313 |
+
value=True,
|
| 1314 |
+
info="Améliore la pertinence"
|
| 1315 |
+
)
|
| 1316 |
+
|
| 1317 |
+
gr.Markdown("### 📊 Statistiques")
|
| 1318 |
+
gr.Markdown(f"""
|
| 1319 |
+
- **Modèle embedding:** Qwen3-Embedding-4B
|
| 1320 |
+
- **Modèle reranking:** Qwen3-Reranker-4B
|
| 1321 |
+
- **Modèle génération:** Qwen3-4B-Instruct-2507
|
| 1322 |
+
- **Index FAISS:** HNSW optimisé
|
| 1323 |
+
- **Vecteurs:** {rag_system.config.total_vectors:,}
|
| 1324 |
+
""")
|
| 1325 |
+
|
| 1326 |
+
# Interactions
|
| 1327 |
+
def _clear_message():
|
| 1328 |
+
return ""
|
| 1329 |
+
|
| 1330 |
+
def _clear_chat():
|
| 1331 |
+
return []
|
| 1332 |
+
|
| 1333 |
+
# Envoi par Entrée
|
| 1334 |
+
msg.submit(
|
| 1335 |
+
chat_with_generic_rag,
|
| 1336 |
+
[msg, chatbot, top_k_slider, reranking_checkbox],
|
| 1337 |
+
chatbot
|
| 1338 |
+
).then(
|
| 1339 |
+
_clear_message,
|
| 1340 |
+
outputs=msg
|
| 1341 |
+
)
|
| 1342 |
+
|
| 1343 |
+
# Envoi par bouton
|
| 1344 |
+
send_btn.click(
|
| 1345 |
+
chat_with_generic_rag,
|
| 1346 |
+
[msg, chatbot, top_k_slider, reranking_checkbox],
|
| 1347 |
+
chatbot
|
| 1348 |
+
).then(
|
| 1349 |
+
_clear_message,
|
| 1350 |
+
outputs=msg
|
| 1351 |
+
)
|
| 1352 |
+
|
| 1353 |
+
# Effacement de la conversation
|
| 1354 |
+
clear_btn.click(_clear_chat, outputs=chatbot)
|
| 1355 |
+
|
| 1356 |
+
return demo
|
| 1357 |
+
|
| 1358 |
+
|
| 1359 |
+
def main():
|
| 1360 |
+
"""Point d'entrée principal."""
|
| 1361 |
+
print("🚀 LocalRAG Step 03 - Interface de chat générique")
|
| 1362 |
+
print("=" * 50)
|
| 1363 |
+
|
| 1364 |
+
# Vérification des dépendances
|
| 1365 |
+
if not _check_dependencies():
|
| 1366 |
+
return 1
|
| 1367 |
+
|
| 1368 |
+
# Initialisation du système RAG
|
| 1369 |
+
global rag_system
|
| 1370 |
+
try:
|
| 1371 |
+
rag_system = _create_rag_system()
|
| 1372 |
+
except Exception as e:
|
| 1373 |
+
print(f"❌ Erreur d'initialisation: {e}")
|
| 1374 |
+
return 1
|
| 1375 |
+
|
| 1376 |
+
# Configuration de l'interface Gradio avec thème Glass
|
| 1377 |
+
with gr.Blocks(
|
| 1378 |
+
title="🤖 LocalRAG Chat Générique",
|
| 1379 |
+
theme=gr.themes.Glass(),
|
| 1380 |
+
) as demo:
|
| 1381 |
+
|
| 1382 |
+
# En-tête simplifié avec composants Gradio natifs
|
| 1383 |
+
with gr.Row():
|
| 1384 |
+
with gr.Column():
|
| 1385 |
+
gr.Markdown("# 🤖 Assistant RAG Générique LocalRAG")
|
| 1386 |
+
|
| 1387 |
+
# Affichage de l'environnement d'exécution
|
| 1388 |
+
env_info = ""
|
| 1389 |
+
if ZEROGPU_AVAILABLE and os.getenv("SPACE_ID"):
|
| 1390 |
+
env_info = "🚀 **Powered by ZeroGPU** - GPU gratuit Hugging Face"
|
| 1391 |
+
elif torch.backends.mps.is_available():
|
| 1392 |
+
env_info = "🍎 **Apple Silicon optimisé** - MPS accelerated"
|
| 1393 |
+
elif torch.cuda.is_available():
|
| 1394 |
+
env_info = f"🐧 **CUDA accelerated** - {torch.cuda.get_device_name()}"
|
| 1395 |
+
else:
|
| 1396 |
+
env_info = "💻 **CPU optimisé** - Traitement local"
|
| 1397 |
+
|
| 1398 |
+
gr.Markdown(f"**Système RAG complet avec modèles Qwen3 de dernière génération**")
|
| 1399 |
+
gr.Markdown(env_info)
|
| 1400 |
+
gr.Markdown(f"🧠 {rag_system.config.embedding_model.split('/')[-1]} • 🎯 Qwen3-Reranker-4B • 💬 Qwen3-4B • ⚡ Recherche en 2 étapes")
|
| 1401 |
+
gr.Markdown(f"📦 Repository: `{rag_system.config.repo_id}` | 📊 Vecteurs: **{rag_system.config.total_vectors:,}**")
|
| 1402 |
+
|
| 1403 |
+
# Interface de chat
|
| 1404 |
+
chatbot = gr.Chatbot(
|
| 1405 |
+
height=500,
|
| 1406 |
+
show_label=False,
|
| 1407 |
+
container=True,
|
| 1408 |
+
show_copy_button=True,
|
| 1409 |
+
autoscroll=True,
|
| 1410 |
+
avatar_images=(None, "🤖"),
|
| 1411 |
+
type="messages"
|
| 1412 |
+
)
|
| 1413 |
+
|
| 1414 |
+
# Zone de saisie
|
| 1415 |
+
with gr.Row():
|
| 1416 |
+
msg = gr.Textbox(
|
| 1417 |
+
placeholder="Posez votre question...",
|
| 1418 |
+
show_label=False,
|
| 1419 |
+
container=False,
|
| 1420 |
+
scale=4
|
| 1421 |
+
)
|
| 1422 |
+
send_btn = gr.Button("📤 Envoyer", variant="primary", scale=1)
|
| 1423 |
+
|
| 1424 |
+
# Panneau de contrôle avancé simplifié
|
| 1425 |
+
with gr.Accordion("🎛️ Contrôles avancés", open=True):
|
| 1426 |
+
with gr.Row():
|
| 1427 |
+
top_k_slider = gr.Slider(
|
| 1428 |
+
minimum=1,
|
| 1429 |
+
maximum=10,
|
| 1430 |
+
value=3,
|
| 1431 |
+
step=1,
|
| 1432 |
+
label="📊 Nombre de documents finaux",
|
| 1433 |
+
info="Documents qui seront utilisés pour générer la réponse"
|
| 1434 |
+
)
|
| 1435 |
+
|
| 1436 |
+
reranking_checkbox = gr.Checkbox(
|
| 1437 |
+
value=True,
|
| 1438 |
+
label="🎯 Activer le reranking Qwen3",
|
| 1439 |
+
info="Améliore la pertinence avec un modèle de reranking spécialisé"
|
| 1440 |
+
)
|
| 1441 |
+
|
| 1442 |
+
# Bouton pour effacer
|
| 1443 |
+
clear_btn = gr.Button("🗑️ Effacer la conversation", variant="secondary", size="lg")
|
| 1444 |
+
|
| 1445 |
+
# Informations en pied de page avec Accordion pour économiser l'espace
|
| 1446 |
+
with gr.Accordion("ℹ️ Informations sur l'architecture", open=False):
|
| 1447 |
+
env_docs = ""
|
| 1448 |
+
if ZEROGPU_AVAILABLE and os.getenv("SPACE_ID"):
|
| 1449 |
+
env_docs = """
|
| 1450 |
+
### 🚀 Optimisations ZeroGPU
|
| 1451 |
+
|
| 1452 |
+
- **Allocation dynamique :** GPU alloué automatiquement pour le reranking et la génération
|
| 1453 |
+
- **NVIDIA H200 :** 70GB VRAM disponible pour les calculs intensifs
|
| 1454 |
+
- **Décorateurs intelligents :** `@spaces.GPU()` pour optimiser l'usage GPU
|
| 1455 |
+
- **Cache optimisé :** Stockage temporaire en `/tmp` pour performances maximales
|
| 1456 |
+
"""
|
| 1457 |
+
elif torch.backends.mps.is_available():
|
| 1458 |
+
env_docs = """
|
| 1459 |
+
### 🍎 Optimisations Apple Silicon
|
| 1460 |
+
|
| 1461 |
+
- **Metal Performance Shaders :** Accélération native Apple
|
| 1462 |
+
- **Index FAISS adapté :** IndexFlatIP pour éviter les segfaults
|
| 1463 |
+
- **Mémoire unifiée :** Partage efficace CPU/GPU
|
| 1464 |
+
- **Float32 :** Précision optimisée pour MPS
|
| 1465 |
+
"""
|
| 1466 |
+
else:
|
| 1467 |
+
env_docs = """
|
| 1468 |
+
### ⚡ Optimisations locales
|
| 1469 |
+
|
| 1470 |
+
- **Multi-plateforme :** Support CPU, CUDA, MPS selon disponibilité
|
| 1471 |
+
- **Flash Attention :** Activé automatiquement sur CUDA
|
| 1472 |
+
- **Gestion mémoire :** Cleanup automatique pour stabilité
|
| 1473 |
+
"""
|
| 1474 |
+
|
| 1475 |
+
gr.Markdown(f"""
|
| 1476 |
+
### 🚀 Architecture LocalRAG Step 03
|
| 1477 |
+
|
| 1478 |
+
- **📥 Step 02 :** Embeddings chargés depuis Hugging Face Hub au format SafeTensors
|
| 1479 |
+
- **🔍 Recherche :** Index FAISS reconstructé pour recherche vectorielle haute performance
|
| 1480 |
+
- **🎯 Reranking :** Qwen3-Reranker-4B pour affiner la sélection des documents
|
| 1481 |
+
- **💬 Génération :** Qwen3-4B-Instruct-2507 pour des réponses contextuelles optimisées
|
| 1482 |
+
{env_docs}
|
| 1483 |
+
### 📊 Lecture des scores
|
| 1484 |
+
|
| 1485 |
+
- **Score Embedding :** Similarité vectorielle initiale (0.0-1.0, plus haut = plus pertinent)
|
| 1486 |
+
- **Score Reranking :** Score de pertinence final après analyse contextuelle
|
| 1487 |
+
- **Changement de rang :** Evolution de la position du document après reranking
|
| 1488 |
+
""")
|
| 1489 |
+
|
| 1490 |
+
# Gestionnaire de likes
|
| 1491 |
+
def like_response(evt: gr.LikeData):
|
| 1492 |
+
print(f"Réaction utilisateur: {'👍' if evt.liked else '👎'} sur le message #{evt.index}")
|
| 1493 |
+
print(f"Contenu: {evt.value[:100]}...")
|
| 1494 |
+
|
| 1495 |
+
chatbot.like(like_response)
|
| 1496 |
+
|
| 1497 |
+
# Envoi par touche Entrée
|
| 1498 |
+
msg.submit(
|
| 1499 |
+
chat_with_generic_rag,
|
| 1500 |
+
[msg, chatbot, top_k_slider, reranking_checkbox],
|
| 1501 |
+
chatbot
|
| 1502 |
+
).then(
|
| 1503 |
+
_clear_message,
|
| 1504 |
+
outputs=msg
|
| 1505 |
+
)
|
| 1506 |
+
|
| 1507 |
+
# Envoi par bouton
|
| 1508 |
+
send_btn.click(
|
| 1509 |
+
chat_with_generic_rag,
|
| 1510 |
+
[msg, chatbot, top_k_slider, reranking_checkbox],
|
| 1511 |
+
chatbot
|
| 1512 |
+
).then(
|
| 1513 |
+
_clear_message,
|
| 1514 |
+
outputs=msg
|
| 1515 |
+
)
|
| 1516 |
+
|
| 1517 |
+
# Effacement de la conversation
|
| 1518 |
+
clear_btn.click(_clear_chat, outputs=chatbot)
|
| 1519 |
+
|
| 1520 |
+
print("🌐 Lancement de l'interface Gradio...")
|
| 1521 |
+
|
| 1522 |
+
# Configuration HTTPS pour Claude Desktop
|
| 1523 |
+
ssl_keyfile = os.getenv("SSL_KEYFILE")
|
| 1524 |
+
ssl_certfile = os.getenv("SSL_CERTFILE")
|
| 1525 |
+
|
| 1526 |
+
if ssl_keyfile and ssl_certfile:
|
| 1527 |
+
print("🔒 Mode HTTPS activé")
|
| 1528 |
+
print("🔗 Serveur MCP : /gradio_api/mcp/sse")
|
| 1529 |
+
|
| 1530 |
+
demo.launch(
|
| 1531 |
+
mcp_server=True, # Toujours activer MCP
|
| 1532 |
+
inbrowser=True,
|
| 1533 |
+
show_error=True,
|
| 1534 |
+
ssl_keyfile=ssl_keyfile,
|
| 1535 |
+
ssl_certfile=ssl_certfile
|
| 1536 |
+
)
|
| 1537 |
+
else:
|
| 1538 |
+
print("🔗 Serveur MCP : /gradio_api/mcp/sse")
|
| 1539 |
+
print("💡 Pour HTTPS : python step03_ssl_generator_optional.py")
|
| 1540 |
+
|
| 1541 |
+
demo.launch(
|
| 1542 |
+
mcp_server=True, # Toujours activer MCP
|
| 1543 |
+
inbrowser=True,
|
| 1544 |
+
show_error=True
|
| 1545 |
+
)
|
| 1546 |
+
|
| 1547 |
+
print("📋 Outil MCP exposé : ask_rag_question")
|
| 1548 |
+
|
| 1549 |
+
return 0
|
| 1550 |
+
|
| 1551 |
+
|
| 1552 |
+
if __name__ == "__main__":
|
| 1553 |
+
exit(main())
|