Update main.py

main.py

@@ -132,9 +132,9 @@ def get_summarizer():
 
 
 MODEL_CHOICES = [
-    "google/flan-t5-small",  # Modèle de réponse généraliste
-    "deepset/roberta-base-squad2",  # Spécialisé en Q/R
-    "cmarkea/flan-t5-base-fr"  # Pour le français
+    "cmarkea/flan-t5-base-fr",  # Best for French
+    "bigscience/bloomz-560m",   # Good multilingual
+    "google/flan-t5-small"      # Fallback
 ]
 
 qa_pipeline = None
@@ -874,7 +874,7 @@ async def question_answering(
     file: Optional[UploadFile] = File(None)
 ):
     if qa_pipeline is None:
-        raise HTTPException(503, detail="QA system unavailable")
+        raise HTTPException(503, detail="Système indisponible")
 
     try:
         # Process file if provided
@@ -882,66 +882,65 @@ async def question_answering(
         if file:
             _, content = await process_uploaded_file(file)
             full_text = extract_text(content, file.filename.split('.')[-1])
-            context = re.sub(r'\s+', ' ', full_text).strip()[:2000]  # Clean and limit context
+            context = re.sub(r'\s+', ' ', full_text).strip()[:1500]  # Clean and limit context
 
-        # Special handling for theme questions
-        theme_keywords = ["thème", "theme", "sujet principal", "quoi le sujet", "de quoi ça parle"]
+        # Special handling for theme detection
+        theme_keywords = ["thème", "theme", "sujet principal", "quoi le sujet"]
         if any(kw in question.lower() for kw in theme_keywords):
             if not context:
                 return {
                     "question": question,
-                    "answer": "Aucun document fourni pour déterminer le thème",
+                    "answer": "Veuillez fournir un document pour identifier le thème",
                     "context_used": False
                 }
             
-            # Special prompt for theme detection
+            # Optimized theme extraction prompt
             theme_prompt = (
-                "Extrait le thème principal en 1-2 phrases en français à partir de ce texte. "
-                "Sois concis et précis. Texte:\n" + context[:1500]
+                "Extrait uniquement le thème principal en une phrase concise en français. "
+                "Ne donne pas d'exemples ou de détails. Texte:\n" + context
             )
             
             theme_result = qa_pipeline(
                 theme_prompt,
-                max_length=150,
-                num_beams=2,
-                temperature=0.3,  # Lower temperature for more focused answers
-                repetition_penalty=2.5
+                max_length=50,  # Very short for single-sentence answers
+                num_beams=1,    # More deterministic
+                temperature=0.1, # Minimal creativity
+                repetition_penalty=3.0,
+                no_repeat_ngram_size=2
             )
             
+            # Post-processing cleanup
+            clean_answer = theme_result[0]["generated_text"].split(".", 1)[0] + "."
+            clean_answer = re.sub(r"^(Le|La)\s+", "", clean_answer)  # Remove articles
+            
             return {
                 "question": question,
-                "answer": theme_result[0]["generated_text"],
+                "answer": clean_answer,
                 "model": current_model,
                 "context_used": True
             }
 
         # Standard QA handling
-        input_text = f"Réponds en français à: {question}"
+        input_text = f"Réponds brièvement en français à: {question}"
         if context:
-            input_text += f" en utilisant ce contexte: {context[:2000]}"
+            input_text += f" en utilisant uniquement ceci: {context[:1000]}"
         
         result = qa_pipeline(
             input_text,
-            max_length=150,
-            num_beams=3,
-            temperature=0.7,
-            repetition_penalty=2.0
+            max_length=100,
+            num_beams=2,
+            temperature=0.3
         )
 
-        # Post-process answer
-        answer = result[0]["generated_text"]
-        if answer.lower().startswith(("question:", "réponse:")):
-            answer = answer.split(":", 1)[1].strip()
-        
         return {
             "question": question,
-            "answer": answer,
+            "answer": result[0]["generated_text"],
             "model": current_model,
             "context_used": context is not None
         }
 
     except Exception as e:
-        logger.error(f"Error: {str(e)}")
+        logger.error(f"Erreur: {str(e)}")
         raise HTTPException(500, "Erreur de traitement")