User Comments

Keep in mind that although MATCH() AGAINST() is case-insensitive, it also is basically **accent-insensitive**. In other words, if you do not want _mangé_ to match with _mange_ (this example is in French), you have no choice but to use the BOOLEAN MODE with the double quote operator. This is the only way that MATCH() AGAINST() will make accent-sensitive matches.

E.g.:

SELECT * FROM quotes_table WHERE MATCH (quote) AGAINST ('"mangé"' IN BOOLEAN MODE)

For multiword searches:

SELECT * FROM quotes_table MATCH (quote) AGAINST ('"mangé" "pensé"' IN BOOLEAN MODE)

SELECT * FROM quotes_table MATCH (quote) AGAINST ('+"mangé" +"pensé"' IN BOOLEAN MODE)

Although the double quotes are intended to enable phrase searching, just like any web search engine for example, you can also use them to signify single words where accents and other diacritics matter.

The only drawback to this method seems to be that the asterisk operator is mutually exclusive with the double quote. Or I just haven't been able to combine both effectively.

Be careful with the phrase search when short words are involved!
Words that are shorter than the minimum word length (by default, words with up to 3 characters) are sometimes taken into consideration when you search for phrases, but sometimes not!

Example 1:
A search for the phrase "the creation" will find all records that really contain this phrase, and only those. So, a record containing only "la creation du monde", even without the accent aigu on the e in creation, won't be found. This is just fine and what one would expect.

Example 2: A search for the phrase "let it be" won't find any record, not even records containing something like "The Beatles: Let It Be". According to the MySQL team, this is not a bug.
I personally find it very counterintuitive to sometimes take short words into consideration for phrase searches, but only if there is at least one properly long word in the search phrase.

For those of you who interface MySQL with PHP and wonder what the problem is with getting "exact phrases" working properly, here's the way to go.

$query= "SELECT code, category, header, date FROM articles WHERE MATCH (text,header,summary) AGAINST ('" . stripslashes (str_replace (""", "\"", ($_POST['keywords']))) . "' IN BOOLEAN MODE)";

PHP, or some setups or with some browsers, convert double quotes from POST data to their HTML-equivalents even without being asked to do that. The above will fix the issue. Stripslashes() is optional, I prefer to keep it in to keep things looking clean, though the \" doesn't seem to break the boolean literal search.

It's also possible to create a prioritized boolean query with the following SQL:

SELECT id, text, MATCH (text) AGAINST ('word1 word2 word3' in boolean mode)
AS score FROM table1
WHERE MATCH (text) AGAINST ('word1 word2 word3' in boolean mode) order by score desc;

According to the last comment by Rob Thorpe it's possible to prioritize the boolean query like so:

SELECT id, text, MATCH (text) AGAINST ('word1 word2 word3' in boolean mode)
AS score FROM table1
WHERE MATCH (text) AGAINST ('word1 word2 word3' in boolean mode) order by score desc;

I tried this and it failed to work, i.e. all the scores turned out to be '1' even though the number of matches differed from record to record.

Boolean searching has two deficiencies: 1) results are not sorted by relevance and; 2) no method by which to weigh certain columns. There is a way around both of these problems. For example, if I have a table of articles and want to weigh the title more heavily than the text, I can do the following:

SELECT *, ( (1.3 * (MATCH(title) AGAINST ('+term +term2' IN BOOLEAN MODE))) + (0.6 * (MATCH(text) AGAINST ('+term +term2' IN BOOLEAN MODE))) ) AS relevance FROM [table_name] WHERE ( MATCH(title,text) AGAINST ('+term +term2' IN BOOLEAN MODE) ) HAVING relevance > 0 ORDER BY relevance DESC;

Here we artificially manipulate the relevancy score to give title more weight by multiplying by the constant 1.3. In the above query, it doesn't seem to matter whether I have 3 fulltext indexes or just one comprising the title and text columns. From my testing, the results appear to be the same.

In response to the note above Posted by Adam George on December 13 2004 7:32pm:

In my tests it would seem that the score returned is an integer equal to the number of words matched. So if you match on 3 words the scores will range from 1 to 3. If you match only on one word, or only one word is matched in any document, then the scores would all be 1.

In response to Joe Laffey and Adam George:
To enhance sorting of the results in boolean mode you can do the following:

SELECT id, text, MATCH (text) AGAINST ('word1 word2 word3')
AS score FROM table1
WHERE MATCH (text) AGAINST ('+word1 +word2 +word3' in boolean mode) order by score desc;

Using the first MATCH() we get the score in non-boolean search mode (more distinctive). The second MATCH() ensures we really get back only the results we want (with all 3 words). If you want to do 'any of the words' search only, it's better to use non-boolean search instead (unless you are using boolean in order to get rid of 50% treshold limit).

I'm not sure why MATCH/AGAINST uses a different scoring method when in boolean mode and when it's not. As stated above, if searching 3 terms in boolean mode, the score will be between 1 and 3 (integer). However, if not in boolean mode, the score is a floating point value.

It seems that non-boolean mode returns a "real" relevancy (based on how often each term was found I presume). In boolean mode it only returns how many terms were found. To me, this is not really relevancy. For instance, if searching on 2 terms, one result may have 20 occurrences of each term and another may have only 1 occurrence of each word, yet they will both return "2" as their relevance.

This seems to work well for ranking relevance in boolean queries:

select products_id,match(products_model) against ('printer' ) as Relevance

from products

where match(products_model) against ('+"printer"' in boolean mode)

Hi,
if You want to combine the phrase search with the *, simply search for the words separately and apply a having clause.

For example:
If You wish to find `bird cathcing`, `bird cathcer`, `bird cathers`

try this

SELECT column, MATCH( column ) AGAINST ('bird catch' IN BOOLEAN MODE) AS rank
FROM mytable
WHERE MATCH( column ) AGAINST ('bird catch' IN BOOLEAN MODE) > 0
HAVING column LIKE '%bird catch%'
ORDER BY rank DESC

I know the HAVING clause is slow, but it is only allpied to the rows that match the search criteira.

Have'n tested on big tables, but I have a feeling it works just fine.

This works for me so I get a score and the benefits of a boolean search. However, it's doing two different fulltext searches so it may slow things down a lot:

SELECT *, MATCH(post_content, post_title) AGAINST('string') AS `score` FROM posts MATCH(post_content, post_title) AGAINST('string' IN BOOLEAN MODE) ORDER BY score DESC LIMIT 10

The boolean mode after the from statement automatically pulls out the 0's and then the Match statement in the SELECT clause allows me to get the relevance score so I can sort by it.

Following on from Robert's comment, I've been playing around using his technique trying to get 'exact' and 'any word' searches to work. I've been having trouble with plurals, e.g. searching for 'anchor bolts' when the fulltext index includes 'anchors' and 'bolt'. This can be solved with the boolean part of the Match statement by trimming and searching for +anchor* +bolt*, but the non-boolean part of the Match statement for relevance is still a problem returning zeros in some instances. My 'fix' for this is to search for 'anchor anchors bolt bolts' in the non-boolean part (fairly easy to program in php), which seems to solve the problem. Any easier solutions such as like '%anchor%'?

I needed to be able to take a user search that might have words in it that are less then the min word length and return only results that contained all words.

What I did was break the query up into two sets of words, one set contained all the words that were >= ft_min_word_len, the other set contained all the shorter words. I did a fulltext search on the words that met the fulltext search length requirement and augmented it with an AND clause that used LIKE '%<shortword1>%'. MySQL uses the fulltext index to narrow down the results and then applies the LIKE conditions, so it stays fast.

This way you don't need to change your ft_min_word_len to a smaller number, which will make your indexes bigger.

Here's the regex I used to break things up:
\b(\w{4,})\b
where ft_min_word_len = 4, that will match all words of 4 or more letters.

Example
Search for the string "axle hub nut" will result in:
WHERE MATCH (col) AGAINST ("+axle*" IN BOOLEAN MODE) AND col LIKE '%hub%' AND col LIKE '%nut%'

Fulltext boolean mode search returning (relevance):
SELECT MATCH (x) AGAINST ('word1 word2' IN BOOLEAN MODE) AS relevance
The returned relevance is 0,1 or 2. 0 = no match, 1 = one of the words is matched (word1 or word2), 2 = both word are matched.
When using eg. '+word1 word2 word3' it returns a floating point number (double) which is the relevance (all returned rows contain word1, some are more relevant (they return word2 and word3, others less..they contain word1 and (word2 or word3).. others even less..containing only word1 (relevance = 1).

Markus Loponen, awesome tip about the PHP statement, I was wondering why the quotes didn't seem to do the trick. Thanks for sharing!

Here's a query I use to return relevance-based data while still using boolean mode searches, and using weights for different columns, and even weights for specific results:

SELECT id,
store,
name,
(((MATCH(name)
AGAINST(?) * 1.2) +
MATCH(description, keywords)
AGAINST (?) +
((MATCH(creators)
AGAINST (?) * 1.2)) / 3) +
(((store IN ('xfx','w3d')) * .12) * ((store IN ('iv', '3da', 'vp')) * .1))) *
MATCH(description,keywords,requirements,creators,name)
AGAINST(? IN BOOLEAN MODE)
AS sort_rel
FROM prod_text
WHERE MATCH(description,keywords,requirements,creators,name)
AGAINST(? IN BOOLEAN MODE)
AND store != 'pp'
HAVING sort_rel > 0.2
ORDER BY sort_rel DESC

All placeholders (?s) take the search terms entered by the user in this case.

For the record, I also like to use three letter (and smaller) words, but I also understand the need to not index all of them (it would simply be absurd).

Instead of enabling a smaller size, what I do is this:

I preprocess against a list of three letter words I want to be matchable. For any row I find, I insert a different string into the indexed text the same number of times (this column need not actually be the one holding returned text mind you, it can be a duplicate column used only for indexing and searching).

When a user searches, I preprocess his or her search query in the same way, replacing any of the words I want matchable with the longer string.

For instance, if I wanted the search to be able to match 'GUN' I would search (using LIKE) for all rows matching and get back the results into a perlscript:

my $check_for_tlw_q = <<"EOF";
SELECT id, fulltext_indexed
FROM some_table
WHERE fulltext_displayed RLIKE '(^|[[:blank:]])GUN([[:blank:]]|$)'
AND fulltext_indexed NOT RLIKE '(^|[[:blank:]])TLWGUNTLW([[:blank:]]|$)'
EOF
my $check_for_tlw = $dbh->prepare($check_for_tlw_q);

my $update_tlw_q = <<<"EOF";
UPDATE some_table
SET fulltext_indexed = ?
WHERE id = ?
EOF
my $update_tlw = $dbh->prepare($update_tlw_q);

$check_for_tlw->execute;
while (my $row = $check_for_tlw->fetchrow_hashref) {
$row->{fulltext_indexed} =~ s/\bgun\b/TLWGUNTLW/igsm;
$update_tlw->execute($row->{fulltext_indexed}, $row->{id});
}

Then when a user searches for 'machine gun' the search is substituted to be 'machine TLWGUNTLW' behind the scenes. This has the effect of making 'gun' matchable without enabling words like 'the' and 'and' and 'les' and 'how' and 'brb' and thus not bogging down the index except where necessary.

Using > < can give some weight but if you have more than two words to match against like match(..) against('>frist second third') and you want to give weight, seems like there is no way. adding '>' will give you the higher weight to 'first' in this case but how to give second and third.. weight is not clear.

This is impportant specially if you have more than two fields to match against, field artist, field album, field producer, you want to list with artist then album then producer.

Note that you can use myisam_ftdump utility to dump the fulltext index.

Another possibility for returning results based on relevance is to construct your search table with a score/value already built in. This is not a perfect solution by any means but will certainly return results much quicker with extremely large tables and help to sort the results a little better.

1. Build your search table

CREATE TABLE my_search
(
id int(10),
category smallint(5),
score smallint(5),
text_data text
)
ENGINE=MyISAM
DEFAULT CHARSET=latin1
ROW_FORMAT=COMPRESSED;

2. Fill the table with data

Fill the search table with data giving a greater score value, to more important data.

For example with a product database:

Insert the product names, make and model into the "text_data" column and give it a score of 5 (or similarly higher score). Insert the product specification and give this a score of 3 (medium value) and then insert the product description and any other related text with a score of 1 (low value).

Therefore, in this example, we insert 3 rows for the three items of data with differing scores relating to the data's importance.

3. Create indexes

CREATE FULLTEXT INDEX idx_1 ON my_search (text_data);
CREATE INDEX idx_2 ON my_search (id);

You may wish to create indexes for the other columns, such as category, if they are included in your proposed SELECT statement.

4. Execute your SELECT

The following Select statement will return your search results sorted with matches against product name towards the top. If the match is against product name *and* specification then this has an even higher score (relevance) and appears higher in the list.

SELECT id, sum(score)
FROM my_search
WHERE MATCH(text_data) AGAINST ('search for this text' IN BOOLEAN MODE)
GROUP BY id
ORDER BY sum(score) DESC;

--

As I said at the beginning, this is not a perfectly accurate list by *real* computed relevance. However, this may provide help in some circumstances with extremely large databases (mine has 2.2 million rows).

Add your own comment.