Site-wide Search

part of the ArsDigita Community System by Philip Greenspun, Jin Choi, and Michael Bryzek
Note: this facility requires interMedia version 8.1.5.1, which you can upgrade to via a patch, or Oracle 8.1.6.

The Big Idea

This is a system for using a full-text search engine bundled with the RDBMS, e.g., Oracle Intermedia, to permit site-wide searches with results ranked by descending relevance. So, for example, bulletin board postings and classified ads are searched at the same time as static documents.

Under the Hood

We define a site_wide_index table which ties together rows from different tables to a single interMedia index column. We use the user datastore feature of interMedia which gives us flexibility in how we index various tables, and saves us from having to copy text we wish to index.

Deeper Under the Hood

The main action is

create table site_wide_index (
	table_name	     	varchar(30) not null,
	the_key		     	varchar(700) not null,
	one_line_description 	varchar(4000) not null,
	datastore		char(1) not null, -- place holder for datastore column
	primary key (table_name, the_key)
);
We store the table name from which the content came originally. We also keep a key from that table. This will typically be an integer but it might be character data (e.g., the msg_id from bboard). In some cases, it might be a delimited composite key (it is up to the application code to pull this out and use it intelligently). Note that Oracle automatically coerces from integer to varchar and therefore we don't have to do anything fancy to make queries work in the typical case (where the primary key of the fundamental table is an integer).

We must have the primary key constraint on this index because Oracle Intermedia can't deal with any table that doesn't have a primary key.

We need one_line_description so that we can display a nice list of query results to users who do a site-wide search. If we didn't keep this, we'd have to outer join with every possible table that contributes to the index or laboriously look up every row with a PL/SQL procedure. And even then we'd need to keep some record of, for each table, what was the best way to get a one-line description.

For consistent site-wide user interface, we keep a table of what the table names in the index should be called as far as sections go. For example, the table name for the /bboard module is "bboard" but a publisher might wish this presented in a search form as "Discussion Forums". This is done by entering rows in the table_acs_properties table, described in the user profiling documentation. We reproduce the table definition here:


create table table_acs_properties (
             table_name      varchar(30) primary key,
             section_name    varchar(100) not null,
             user_url_stub   varchar(200) not null,
             admin_url_stub  varchar(200) not null
);
The user_url_stub column holds the url necessary to direct the user from the search results page to the page with the content. For example, for the discussion forum module, this column would contain "/bboard/redirect-for-sws?msg_id=". The search results page will append the value of the_key column to this URL stub.

Triggers that fill the table

Here's some example PL/SQL code to keep the site-wide index table updated for static pages. It is complicated by the need to observe the index_p flag, which is used to mark pages which should not be indexed.

Note that the datastore column is just a placeholder; its value is unimportant (we just happen to use the letter 'a'). Updating this column causes the index to be rebuilt for a row by calling the user datastore procedure (see next section).


create or replace trigger static_pages_sws_insert_tr
  after insert on static_pages for each row
BEGIN
  IF :NEW.index_p = 't' THEN
     -- we have to create a new row in the index table for this row.
     insert into site_wide_index (table_name, the_key, one_line_description, datastore)
       values ('static_pages', :new.page_id, :new.page_title, 'a');
  END IF;
END;
/
show errors
  

CREATE OR replace trigger static_pages_sws_update_tr
  after UPDATE ON static_pages FOR each row
BEGIN
  IF :old.index_p = 'f' AND :NEW.index_p = 't' THEN
     insert into site_wide_index (table_name, the_key, one_line_description, datastore)
       values ('static_pages', :new.page_id, :new.page_title, 'a');
  ELSIF :old.index_p = 't' AND :NEW.index_p = 'f' THEN
     DELETE FROM site_wide_index
       WHERE table_name = 'static_pages'
       AND the_key = :old.page_id;
  ELSIF :NEW.index_p = 't' THEN
     update site_wide_index 
       set the_key = :new.page_id, one_line_description = nvl(:new.page_title, '(no title)'), datastore = 'a'
       where the_key = :old.page_id
       and table_name = 'static_pages';
  END IF;
end;
/
show errors


CREATE OR replace trigger static_pages_sws_delete_tr
  after DELETE ON static_pages FOR each row
BEGIN
  IF :old.index_p = 't' THEN
     DELETE FROM site_wide_index
       WHERE the_key = :old.page_id
       AND table_name = 'static_pages';
  END IF;
END;
/
show errors


User Datastore Procedure

This is the heart of the site wide index. This procedure needs to know about every section of the ACS which intends to use the site-wide search. This example indexes bboard and static_pages.

This procedure is run to gather the text that is to be indexed for any row of the site_wide_index table. Its arguments are the rowid of the row that it is to retrieve the content for, and a temporary clob in which the results are to be stored.

Different sections can be indexed in different ways. Note that the bboard section indexes entire threads into a single entry in the site wide index, instead of indexing each message as its own entry.

The user datastore procedure must be loaded as the ctxsys user, and must be granted select access on the tables you want to index from. You must grant execute permissions on the user datastore procedure to the user who will be using it to create the index.


CREATE OR replace procedure sws_user_datastore_proc ( rid IN ROWID, tlob IN OUT nocopy clob )
IS
   v_table_name  VARCHAR(30);
   v_primary_key VARCHAR(700);
   v_one_line VARCHAR(700);
   v_static_pages_row photonet.static_pages%ROWTYPE;
   cursor bboard_cursor(v_msg_id CHAR) IS
     SELECT one_line, message 
       FROM photonet.bboard
       WHERE sort_key LIKE v_msg_id || '%';
BEGIN
     -- get various info on table and columns to index
   SELECT table_name, the_key, one_line_description
     INTO v_table_name, v_primary_key, v_one_line
     FROM photonet.site_wide_index
     WHERE rid = site_wide_index.ROWID;
   
   -- clean out the clob we're going to stuff
   dbms_lob.trim(tlob, 0);
   
   -- handle different sections
   IF v_table_name = 'bboard' THEN

      -- Get data from every message in the thread.
      FOR bboard_record IN bboard_cursor(v_primary_key) LOOP
	 IF bboard_record.one_line IS NOT NULL THEN
	    dbms_lob.writeappend(tlob, length(bboard_record.one_line) + 1, bboard_record.one_line || ' ');
	 END IF;
	 dbms_lob.append(tlob, bboard_record.message);
      END LOOP;
   ELSIF v_table_name = 'static_pages' THEN
      SELECT * INTO v_static_pages_row
	FROM photonet.static_pages
	WHERE page_id = v_primary_key;
      
      IF v_static_pages_row.page_title IS NOT NULL THEN
         dbms_lob.writeappend(tlob, length(v_static_pages_row.page_title) + 1, v_static_pages_row.page_title || ' ');
      END IF;
      dbms_lob.append(tlob, v_static_pages_row.PAGE_BODY);
   END IF;
END;
/
show errors

grant execute on sws_user_datastore_proc to photonet;


Querying

If you just want to find the most relevant documents across the entire site:

select 
  score(10) as the_score,
  the_key,
  one_line_description,
  map.user_url_stub 
from site_wide_index swi, table_acs_properties map
where swi.table_name = map.table_name
and contains(indexed_stuff,'about($user_entered_query)',10) > 0 
order by score(10) desc
Within the discussion forums, you wouldn't bother to join with table_acs_properties since you don't need the URL stub and you know what the section is called. But you'd probably want to join with the bboard table in order to restrict to a particular topic, e.g.,

select 
  score(10) as the_score,
  msg_id,
  one-line,
  posting_time,
  topic
from site_wide_index swi, bboard
where swi.the_key = bboard.msg_id 
and swi.table_name = 'bboard'
and contains(indexed_stuff,'about($user_entered_query)',10) > 0
order by score(10) desc

The about() in the query specifies a theme query, which takes the words of the query to be "themes" to search for. Oracle has had geniuses locked up for the last several years dividing up the English language into a taxonomy of semantic topics. interMedia will decide which of these themes your documents are most relevant to, and match those up to the words in your query. Leaving out the about() gives you a simple word query, which is more like what you'd expect.

In practice, feeding a user entered query directly to interMedia turns out to be very bad, because it is expecting queries to be specified using their little query language. Any syntactical errors in defining a query causes interMedia to cough up an error. This is very easy to do, since various punctuation characters and words like "and" are special in this query language. We use the PL/SQL function im_convert to massage the user input into a form which is safe to pass to interMedia, and which performs a combination of text search and theme search to try to bring the most relevant documents.

im_convert() cannot be called directly from within contains(), so we must use a separate database query to convert the user search string. This may be replaced at a later date into a Tcl procedure, to make modifications easier.

Tcl processing of the results

Unlike Verity and PLS, ConText doesn't have a good way to refer to previous searches and say "give me results 100 through 200 from that last search". Combined with the stateless nature of HTTP, this makes it hard to implement the kinds of search user interfaces prevalent at most sites (notably AltaVista). Personally I'm not sorry. I've always hated sites like those. My personal theory is that the user should get about 25 highly relevant documents back from a search. Additional documents shouldn't be offered unless they are nearly as relevant as the most relevant. Here's an example of the Tcl code necessary to implement this:

set counter 0 
while {[ns_db getrow $db $selection]} {
    set_variables_after_query
    incr counter
    if { ![info exists max_score] } {
	# first iteration, this is the highest score
	set max_score $the_score
    }
    if { ($counter > 25) && ($the_score < [expr 0.3 * $max_score] ) } {
	# we've gotten more than 25 rows AND our relevance score
	# is down to 30% of what the maximally relevant row was
	ns_db flush $db
	break
    }
    if { ($counter > 50) && ($the_score < [expr 0.5 * $max_score] ) } {
	# take a tougher look
	ns_db flush $db
	break
    }
    if { ($counter > 100) && ($the_score < [expr 0.8 * $max_score] ) } {
	# take a tougher look yet
	ns_db flush $db
	break
    }
...
}

The heuristic cutoff in the above procedure has been packaged into the Tcl procedure ad_search_results_cutoff.

For the user who has a PhD in computer science

Suppose that the user doesn't like the rough results that come back from a phrase relevancy query in ConText. We should give the user an opportunity to go into an "advanced search" mode where they can exploit the full power of the ConText query language.

"More Like This"

interMedia doesn't have a built-in facility for doing query by example ("give me more documents like this one"), but we can use theme searches to build a reasonable facsimile. The Tcl procedure ad_search_qbe_get_themes takes a table name and key from the site wide index table and returns the top five themes associated with that document. Those themes can be used in an about() search to find documents about the same themes.

In practice, the themes returned by interMedia for a document are fairly general, and occasionally irrelevant (including things like HTML tags). So we throw in the one_line_description as part of the query, which seems to do a better job of bringing up relevant results. See /search/query-by-example for an example of query by example.

Using a Separate Search Server

You may decide you want a separate server to serve up the search queries, perhaps for performance reasons, or perhaps because Oracle hasn't made available the patch you need to run Oracle on a machine with more than 10 IP addresses on the architecture your main server is on. The two parameters, BounceQueriesTo and BounceResultsTo in the site-wide-search section of your parameters file allows you to do this. BounceQueriesTo should be set on your main server, and BounceResultsTo set on the search server. This will bounce queries and results back and forth for all search pages.

Keeping the index in sync

If you update a table with a ConText index.... nothing happens. The index doesn't get updated, unlike any other kind of index in the SQL world. There are a few ways to update the index: The method used to update the index is independent of the rest of the site-wide search module. At ArsDigita, we most often use ad-context-server because it is easy to configure and it is robust enough to prevent deadlocks while updating interMedia indexes.

If you want to manually update the index sws_ctx_index, you can execute:


alter index sws_ctx_index rebuild online parameters('sync');
Be careful that no other process is updating the index at the same time or your index may become unusable.

Optimizing interMedia indexes

Another thing that you might have to do periodically is

alter index sws_ctx_index rebuild online 
parameters('optimize full maxtime 60');
This gives Oracle one hour (60 minutes) to rebuild an optimized index. I'm not quite sure whether this simply means more efficient for query time or better (more relevant) results for users.

If you care about performance

If you don't want to slow down transactions by building up these clob copies of everything, keep the site_wide_index table on a separate physical disk drive. Also try to put the Intermedia index onto yet another separate physical disk drive. Since every disk drive on a 24x7 server must be mirrored, that means you need four new disk drives to implement this glorious module. If you happen to have two more disk drives, use them for the primary key index of the site_wide_index table.
philg@mit.edu