You've probably already thought through this and know the answer, but just in case...

Is there no alternative design to the one that creates a large file which keeps getting larger, passing the several-GB mark and beyond? It might be more efficient, from a searching standpoint to divide the dataset into records and storing them in a relational database for easy searching capability.

If that's not a possibility, how about at least maintaining fixed-size records or entries in the data file, so that you can seek to specific records within the file quickly, without re-reading it constantly. You could even maintain a separate index file of where "matches" are known to exist.

Of course this is all just speculation, but it seems that if you're re-scanning this file at various intervals, and the file is growing to multi-GB sizes, eventually you'll either need to split it up, or cache the search results to maintain scalability.