Hi there, 

i am looking for the insight on the following problem from people who successfully implemented and went live with the following scenario and may address the issues with verified in practice solutions.  

we have implemented a service agent that must be exclusively grounded in the knowledge base with multiple subagents that share only one custom prompt action (now) that triggers the retriever (custom) with a related search index (vector search, with multilingual e5 large embedding model, on knowledge article object (data cloud). the agent is a service agent built in the new builder. prompt llm model is gemini 2.5 pro. 

the knowledge base (standard sf knowledge solution) contains only 89 articles 9some very long, some shorter) but the information is law related and it is very complicated, it is all in English whereas the agent should support also the conversations in French, Spanish and Portuguese (as of now, we also need to roll out other languages after they stop being beta). 

the subagents instructions and the prompt are also in english. 

the problem we have (as of now there are 2 main levels):  

1. agent wise: when the question (that was previously checked in english and returned a good answer) is asked in other languages  the agent either does not understand the question at all or when it does understand it  the answer comes back but is not satisfactory.  

2. search index/retriever wise: when tested directly in the retriever we see that the scoring for the same chunks is lower significantly for other languages in comparison with the tested same question in english (open question if the scoring indeed passes together with chunks to LLM). to patch this we added keywords to knowledge articles in multiple languages, it helps but does not fix the problem. 

the main proposed solutions (also by salesforce): 

1. translate the knowledge base to all those languages we want the agent to speak (which still leaves open the fact that all the instructions and the prompt are in english).  

2. normalize the query (user input) by translating it first to english and then query it while sending to prompt original language, original question and the translated english question. very simplistic description, understandable there is much more to it.  

it feels that such behaviour must be handled totally out of the box and all proposed solutions are intuitive but feel an overkill for something that just should work. 

is there anything i am missing? any insight will be very much appreciated.  

Thank oyu 

Yuliya