I think you are answering your own question here, @Lacona. Can you really say these results are better if you are trimming away so many (high quality) read sequences?
A few links to help troubleshoot issues with low non-chimeric reads were recently shared in a response here.