.\"------------------------------------------------------------
.\" Id - set Rv,revision, and Dt, Date using rcs-Id tag.
.de Id
.ds Rv \\$3
.ds Dt \\$4
..
.Id $Id$
.\"------------------------------------------------------------
.TH mg 1 \*(Dt CITRI
.\"=====================================================================
.\" Author:
.\"     Nelson H. F. Beebe
.\"     Center for Scientific Computing
.\"     Department of Mathematics
.\"     University of Utah
.\"     Salt Lake City, UT 84112
.\"     USA
.\"     Email: beebe@math.utah.edu (Internet)
.\"=====================================================================
.if t .ds Bi B\s-2IB\s+2T\\h'-0.1667m'\\v'0.20v'E\\v'-0.20v'\\h'-0.125m'X
.if n .ds Bi BibTeX
.if t .ds Te T\\h'-0.1667m'\\v'0.20v'E\\v'-0.20v'\\h'-0.125m'X
.if n .ds Te TeX
.\"=====================================================================
.SH NAME
mg \- full-text inverted index support
.\"=====================================================================
.SH DESCRIPTION
.B mg
is a suite of programs that can be used to create
and query full-text inverted indexes for a
collection of documents.
.PP
An inverted index is essentially a list of
pointers to occurrences of every word in a
collection of documents, so that, for example, in
a collection of Shakespeare's works, one can pose
questions like
.I "How often is a fool mentioned in the plays?"
and
.I "Are Romeo and Juliet ever mentioned in the same sentence?"
While search utilities like the UNIX
.BR grep (1)
family can be used for simple queries, they suffer
from several limitations:
.TP \w'\(bu'u+2n
\(bu
Each invocation requires a separate pass over the
documents, which becomes impossibly slow if the
document collection is very large.
.TP
\(bu
It is difficult to compose Boolean queries like
.IR "Caesar AND Brutus AND NOT Antony" .
.TP
\(bu
They provide no mechanism for ranking matches
according to importance, which is extremely
important when queries are complex and numerous
matches are found.
.TP
\(bu
They provide no mechanism for displaying
surrounding context (although the Free Software
Foundation
implementations for the GNU Project remedy this
with their
.BI \-A " num"
(after)
and
.BI \-B " num"
(before)
switches).
.TP
\(bu
They provide no mechanism for searching for
occurrences in the same paragraph, unless
preprocessing is done on the files to wrap
paragraphs into long lines.  Even that will often
fail, because input buffer sizes may be limited to
a few hundred characters.
.TP
\(bu
They provide no easy way to deal with grammatical
word-ending variants, except by explicit
enumeration, such as searching for
.IR compress ,
.IR compressed ,
.IR compresses ,
.IR compressing ,
and
.IR compression .
.PP
Most computer users have been faced with the
problem of finding information from a large
collection of files, such as electronic mail,
on-line documentation, or source code.  Inverted
indices provide a highly-effective solution to
this problem.
.PP
The
.B mg
software is described in Appendix A of the book
.RS
.nf
Ian H. Witten, Alistair Moffat, and Timothy C. Bell
.I "Managing Gigabytes: Compressing and Indexing Documents and Images"
Van Nostrand Reinhold
1994
xiv + 429 pages
US$54.95
ISBN 0-442-01863-0
Library of Congress catalog number TA1637 .W58 1994
.fi
.RE
.PP
Many of the algorithms implemented in
.B mg
represent significant advances over previous work,
both in speed, and in storage requirements.  On a
fast workstation, in tens of minutes, or a few
hours,
.BR mgbuild (1)
can create an index to
.I all
of the words in hundreds of thousands of documents
occupying hundreds of megabytes, or even more than
a gigabyte, of disk space.
.BR mgquery (1)
can then be used to answer complex queries, with
responses often returned in a second or less.
.B mg
also contains algorithms to deal with images, so
that with a small amount of descriptive text for
each image, it is possible to do searches in
collections of images, and to have retrievals
display the images using a viewer like
.BR xv (1).
.PP
.B mg
can deal with compressed text and image files and
surprisingly, it usually runs faster than it would
if the files were not compressed!  Thus, the
considerable disk space savings possible from
file compression are not lost because of the need
for fast document search and retrieval.
.PP
The Free Software Foundation GNU Project
compression utilities
.BR gzip (1)
and
.BR gunzip (1)
are recommended for general use over older
alternatives, like
.BR compress (1),
because of their speed and high compression
ratios.
.\"=====================================================================
.SH AVAILABILITY
The
.B mg
software can be obtained via anonymous ftp to the
Australian archive host
.I "munnari.oz.au [128.250.1.21]"
from the directory
.IR "/pub/mg" .
.\"=====================================================================
.SH "TYPICAL USE OF mg"
Although
.B mg
consists of more than 20 separate programs, many
of which have complicated command-line options,
take heart: most users require only two or three
of these programs, and nothing more than a
document name on the command line.
.PP
A
.I document
for
.B mg
is a fragment of text suitable for retrieval as a
unit when it is found to contain a requested word,
or words.  In a collection of poetry, a document
might be a stanza, while in a novel, it could be a
paragraph.  In an index of first lines of poems, a
document would likely be just a single line.
.PP
Just what constitutes a document is decided by a
user-modifiable UNIX shell script,
.BR mg_get (1).
The default script provided with the
.B mg
source distribution knows about these named
document collections:
.TP \w'mailfiles'u+2n
.I alice
Lewis Carroll's
.I "Alice in Wonderland"
book.
.TP
.I allfiles
all mail files in the directory tree
.IR $HOME/Mail ,
including all of its nested subdirectories.
.TP
.I mailfiles
individual mail messages in
.IR $HOME/mbox
and
.IR $HOME/.sentmail.
.TP
.I davinci
A small collection of text and images from the
work of Leonardo da Vinci.
.PP
A document collection name is used by
.B mg
as a
.BR csh (1)
.I case
statement selector, and as a subdirectory name in the
.B $MGDATA
directory, or the current directory, if
.B MGDATA
is not defined.
.\"=====================================================================
.SH "EXTENDING mg_get"
This section describes how to extend
.BR mg_get (1)
to handle a new document collection.
.PP
Let us take two examples: all \*(Bi
.I .bib
files, and all \*(Te files, contained in
subdirectories under the login directory.
.PP
For \*(Bi, each bibliographic entry will be
considered a separate document.  In order to
facilitate easy identification of entries, we
shall require them to begin at the start of a
line; the
.BR bibclean (1)
utility can be used to standardize the format of
.I .bib
files, and to validate their string values, so
that this requirement is met.
.PP
For \*(Te, each paragraph will be a separate
document, and we assume that paragraphs are
separated by blank lines.  We assume that files
with extensions
.IR .atx ,
.IR .ltx ,
.IR .stx ,
and
.I .tex
contain input to common \*(Te macro package
variants.
.PP
Make a personal copy of the
.I mg_get
script, using the one in the
.B mg
source distribution
.RI ( mg-1.0/mg/mg_get.sh ),
or the one in the local binary program directory,
at many sites called
.IR /usr/local/bin/mg_get .
.PP
Examination of the
.I mg_get
script shows that each document collection name is
used in a
.BR csh (1)
.I case
statement selector, and that most of work is done
by very simple
.BR awk (1)
programs that extract documents from files.  In
your private copy of the
.I mg_get
file, after the line
.PP
.nf
  breaksw #davinci
.fi
.PP
and before the line
.PP
.nf
  default:
.fi
.PP
insert this new code:
.PP
.nf
  case bibfiles:
# Takes a list of files that contain BibTeX entries, and splits them up
#   by putting ^B after each entry. Assumes that each entry
#   begins with a line '^@'.
  switch ($flag)
    case '-init':
    breaksw

    case '-text':
      find $HOME -name '*.bib' -print | \e
        sort | \e
        xargs -l100 awk \e
          '/^@/&&NR!=1{print "^B"} {print $0} END{print "^B"}'
    breaksw #-text

    case '-cleanup':
    breaksw #-cleanup

  endsw #flag
  breaksw #bibfiles

  case texfiles:
# Takes a list of TeX files and split them up
#   by putting ^B after each paragraph. Assumes that each entry
#   begins with a line '^@'.
  switch ($flag)
    case '-init':
    breaksw

    case '-text':
      find $HOME -name '*x' -print | \e
        egrep '[.]tex$|[.]ltx$|[.]atx$|[.]stx$' | \e
        sort | \e
        xargs -l100 nawk '   /^ *$/ {if (b!=1) printf "^B";b=1} \e
            \e!/^ *$/ {print;b=0} \e
            END {printf "^B"}'
    breaksw #-text

    case '-cleanup':
    breaksw #-cleanup

  endsw #flag
  breaksw #texfiles
.fi
.PP
The ^B characters here are Control-B characters,
.I not
caret-B pairs.
.PP
If you have a large number of \*(Bi or \*(Te
files, it is likely that a list of them would be
too long for the UNIX shell to hold in a single
variable, or on a single command line.  Thus,
instead of storing the output of
.BR find (1)
in a variable, we proceed more cautiously, and
employ it to produce a list of the required files,
then pipe them to
.BR xargs (1),
which in turn passes up to 100 filenames at a time
to
.BR nawk (1)
for document selection.
.PP
Install this modified
.I mg_get
script in your private directory for executable
programs (e.g.
.IR $HOME/bin ),
create a directory
.I $HOME/mgdata
to hold the index, issue a
.B rehash
command if you are using
.BR csh (1)
or
.BR tcsh (1),
ensure that
.I mg_get
occurs in your search path
.I before
any system-wide one (the command
.BI which " mg_get"
will tell you which version will be selected),
then create the inverted indexes by
.BI mgbuild " bibfiles"
and
.B mgbuild
.IR texfiles .
These commands may take several minutes to run if
you have a lot of \*(Bi or \*(Te files, or a large
home directory tree.  Once they are complete, you
can then query the index with the commands
.BI mgquery " bibfiles"
and
.B mgquery
.IR texfiles .
These should respond very rapidly.
.PP
In order to keep your index up-to-date, you should
arrange for it to be recreated automatically and
regularly, probably every night.  You can do this
with
.BR cron (1).
Use the command
.B "crontab \-e"
to edit your
.I crontab
file and add two lines like this:
.PP
.nf
00 04 * * * mgbuild bibfiles >$HOME/mgdata/bibfiles.log 2>&1
15 04 * * * mgbuild texfiles >$HOME/mgdata/texfiles.log 2>&1
.fi
.PP
Save the file and exit the editor.  Now, every
night at 4am and 4:15am,
.BR mgbuild (1)
will reconstruct your inverted indexes, and the
results of the builds will be saved in log files
in your
.I $HOME/mgdata
directory.
.\"=====================================================================
.SH "SEE ALSO"
.na
.BR awk (1),
.BR bibclean (1),
.BR bibtex (1),
.BR compress (1),
.BR csh (1),
.BR grep (1),
.BR gunzip (1),
.BR gzip (1),
.BR mg_compression_dict (1),
.BR mg_fast_comp_dict (1),
.BR mg_get (1),
.BR mg_invf_dict (1),
.BR mg_invf_dump (1),
.BR mg_invf_rebuild (1),
.BR mg_passes (1),
.BR mg_perf_hash_build (1),
.BR mg_text_estimate (1),
.BR mg_weights_build (1),
.BR mgbilevel (1),
.BR mgbuild (1),
.BR mgdictlist (1),
.BR mgfelics (1),
.BR mgquery (1),
.BR mgstat (1),
.BR mgtic (1),
.BR mgticbuild (1),
.BR mgticdump (1),
.BR mgticprune (1),
.BR mgticstat (1),
.BR nawk (1),
.BR tcsh (1),
.BR tex (1),
.BR xargs (1),
.BR xmg (1),
.BR xv (1).
.\"=====================================================================
.\" This is for GNU Emacs file-specific customization:
.\" Local Variables:
.\" fill-column: 50
.\" End: