The implementation in FBClob always uses the connection character set, which is not always correct. It should either use the column character set, or maybe even request the actual blob info (which might be too much overhead though). Inspect the various ways blobs are retrieved and written, and try and fix this (maybe with a backwards compatibility option?).