Update app.py

app.py

@@ -37,12 +37,12 @@ def generate_chat_response(request, llm):
             temperature=request.temperature
         )
         reply = response['choices'][0]['message']['content']
-        return reply
+        return {"response": reply, "literal": user_input}
     except Exception as e:
-        return f"Error: {str(e)}"
+        return {"response": f"Error: {str(e)}", "literal": user_input}
 
 def select_best_response(responses, request):
-    coherent_responses = filter_by_coherence(responses, request)
+    coherent_responses = filter_by_coherence([resp['response'] for resp in responses], request)
     best_response = filter_by_similarity(coherent_responses)
     return best_response
 
@@ -62,7 +62,6 @@ def filter_by_similarity(responses):
 
 @app.post("/generate_chat")
 async def generate_chat(request: ChatRequest):
-    # Ejecutar en ThreadPoolExecutor sin límite explícito de workers
     with ThreadPoolExecutor(max_workers=None) as executor:
         futures = [executor.submit(generate_chat_response, request, llm) for llm in llms]
         responses = []
@@ -70,12 +69,17 @@ async def generate_chat(request: ChatRequest):
             response = future.result()
             responses.append(response)
     
-    if any("Error" in response for response in responses):
-        error_response = next(response for response in responses if "Error" in response)
-        raise HTTPException(status_code=500, detail=error_response)
+    if any("Error" in response['response'] for response in responses):
+        error_response = next(response for response in responses if "Error" in response['response'])
+        raise HTTPException(status_code=500, detail=error_response['response'])
     
-    best_response = select_best_response(responses, request)
-    return {"response": best_response}
+    best_response = select_best_response([resp['response'] for resp in responses], request)
+    
+    return {
+        "best_response": best_response,
+        "all_responses": [resp['response'] for resp in responses],
+        "literal_inputs": [resp['literal'] for resp in responses]
+    }
 
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)