I think halley has it right with building an index using an MD5 checksum of your paragraphs is the way to go.

I though it worth mentioning that you can easily read a file, paragraph by paragraph by setting local $/ = '';

The following script builds an in memory index (filename/byte offset) for every file in a list given on the command line.

It indexed every paragraph in 512 html files in my perl distribution in under 10 seconds and 6MB.

#! perl -slw
use strict;
use G; ## Expands command line wildcards
use Data::Dumper;
use Digest::MD5 qw[ md5_hex ];

local $/ = ''; ## paragraph mode.

my( $pos, %index ) = 0; ## The first para start at offset 0

while( <> ) {
    ## build a HoAoAs, MD5 is the key 
      ## The values are arrays of [ filename, offset ].
    push @{ $index{ md5_hex( $_ ) } }, [ $ARGV, $pos ];
    
    ## Getthe next offset
    $pos = tell ARGV;
    
    ## Back to 0 if we reached the EOF
    $pos = 0 if eof( ARGV );
}

print Dumper \%index;
__END__
C:\Perl\html>p:359522 *.html *\*.html *\*\*.html
Processed 512 files and 5829 paragraphs into 3293 unique signatures.
[download]

By storing the offset, you can seek to the place n the file(s) that have a matching MD5 and verify that the paras are indeed identical. There is a small, but statistically possible chance of collisions.