Compression performance depends very much on the documents. If the document contains a random string pattern, those are much less compressible compared to documents containing textual paragraph, for example.
Could you provide the example document that you tried with? And have you tried this experiment with different document patterns? It’ll be great if you can provide information and scripts to reproduce your tests.
As an aside, I usually use a tool like mgeneratejs to generate dummy example documents in large numbers. For example, I can easily create gigabytes of documents following a pattern using this.
Best regards
Kevin