Update main.py

main.py

@@ -874,71 +874,75 @@ async def question_answering(
     file: Optional[UploadFile] = File(None)
 ):
     if qa_pipeline is None:
-        raise HTTPException(
-            status_code=503,
-            detail={
-                "error": "QA system unavailable",
-                "status": "No working model could be loaded",
-                "recovery_suggestion": "Please try again later"
-            }
-        )
-    
+        raise HTTPException(503, detail="QA system unavailable")
+
     try:
-        # Process input
+        # Process file if provided
         context = None
         if file:
-            try:
-                _, content = await process_uploaded_file(file)
-                context = extract_text(content, file.filename.split('.')[-1])[:1000]  # Smaller context
-            except Exception as e:
-                logger.error(f"File processing failed: {str(e)}")
-                raise HTTPException(422, detail=f"File error: {str(e)}")
-
-        # Generate response - MODIFIED PROMPT HERE
-        try:
-            input_text = f"Réponds à cette question: {question}"  # Changed prompt
-            if context:
-                input_text += f" en utilisant ce contexte: {context[:1000]}"  # Added context
-                
-            result = qa_pipeline(
-                input_text,
-                max_length=100,
-                num_beams=2,
-                temperature=0.7,
-                repetition_penalty=2.0
+            _, 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
+
+        # Special handling for theme questions
+        theme_keywords = ["thème", "theme", "sujet principal", "quoi le sujet", "de quoi ça parle"]
+        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",
+                    "context_used": False
+                }
+            
+            # Special prompt for theme detection
+            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]
             )
             
-            # ADDED CHECK FOR QUESTION REFORMULATION
-            if "question:" in result[0]["generated_text"].lower():
-                result[0]["generated_text"] = (
-                    "[Note: This is a question-generation model. For better answers, "
-                    "please use a QA model like google/flan-t5-base].\n"
-                    f"Reformulated question: {result[0]['generated_text']}"
-                )
+            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
+            )
             
             return {
                 "question": question,
-                "answer": result[0]["generated_text"],
+                "answer": theme_result[0]["generated_text"],
                 "model": current_model,
-                "context_used": context is not None
+                "context_used": True
             }
-            
-        except Exception as e:
-            logger.error(f"Generation failed: {str(e)}")
-            raise HTTPException(
-                status_code=500,
-                detail={
-                    "error": "Answer generation failed",
-                    "model": current_model,
-                    "suggestion": "Try a simpler question or smaller document"
-                }
-            )
-            
-    except HTTPException:
-        raise
+
+        # Standard QA handling
+        input_text = f"Réponds en français à: {question}"
+        if context:
+            input_text += f" en utilisant ce contexte: {context[:2000]}"
+        
+        result = qa_pipeline(
+            input_text,
+            max_length=150,
+            num_beams=3,
+            temperature=0.7,
+            repetition_penalty=2.0
+        )
+
+        # 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,
+            "model": current_model,
+            "context_used": context is not None
+        }
+
     except Exception as e:
-        logger.critical(f"Unexpected error: {str(e)}")
-        raise HTTPException(500, "Internal server error")
+        logger.error(f"Error: {str(e)}")
+        raise HTTPException(500, "Erreur de traitement")