The response from the playground give accurate and wanted response, but using the api the accuracy is poor. I am using the same assistant. I followed the process of how to implement the assistant from the openai documentation. Not sure how best to troubleshoot this/ is there an obvious fix?
I’ve also been experiencing the same issue 
see if in the thread you are using v2, when starting thread send the vector id,
“tool_resources” =>
[
“file_search” => [
“vector_store_ids” => [vector_id]
]
]