Re: FTS uses "tsquery" directly in the query

On Mon, 25 Jan 2010 07:19:59 -0800 (PST)
xu fei <autofei(at)yahoo(dot)com> wrote:

> Hi, Oleg Bartunov:
> First thanks for your quick replay. Could you explain it a little
> more on "it's general limitation/feature"? I just confuse that
> to_tsquery('item') function will return a tsquery type which is
> same as 'item'::tsquery, to my understanding. Let me explain what
> I want:First Step: extract top K tokensI have a table with a
> column as tsvector type. Some records in this column are too big,
> which contain hundreds tokens. I just want the top K tokens based
> on the frequency, for example top 5. I am not sure there is a
> direct way to get such kind top K tokens. I just read them out in
> Java and count frequency for each token and sort them. Second
> Step: generate queryNow I will use these tokens to construct a
> query to search other vectors in the same table. I can not
> directly use to_tsquery() due to two reasons: 1) The default logic
> operator in to_tsquery() is "&" but what I need it "|". 2) Since
> the tokens are from tsvector, they are already normalized. If I
> use to_tsquery() again, they will be normalized again! For
> example, “course” -> “cours” -> “cour”. So I just concatenate the
> top K tokens with “|” and directly use "::tsquery ".
> Unfortunately, as you say "it's general limitation/feature”, I can
> not do that. I checked your manual “Full-Text Search
> in PostgreSQL A Gentle Introduction”, but could not figure out
> how. So is it possible to implement what I want in FTS? If so,
> how? Thank! Xu --- On Sun, 1/24/10, Oleg Bartunov

You're trying to solve a similar problems than mine.
I'd like to build up a näive similar text search.
I don't have the "length" problem still I'd like to avoid to
tokenize/lexize a text twice to build up a tsquery.
I've weighted tsvectors stored in a column and once I pick up one
I'd like to look for similar ones in the same column.

There are thousands way to measure text similarity (and Oleg pointed
me to some), still ts_rank should be "good enough for me".

I've very short text so I can't use & on the whole tsvector
otherwise there will be very high chances to find just one match.

As you suggested I could pick up a subset of "important"[1] lexemes
in the tsvector and build up an "&"ed tsquery with them.

Still at least in my case, since I'm dealing with very short texts,
this still looks too risky (just 1 match). Considering that I'm
using weighted tsvectors it seems that "|" and picking up the ones
with the best rank could be a way to go.

But as you've noted there is no function that turns a tsvector in a
tsquery (including weight possibly) and give you the choice to use
"|".

Well... I'm trying to write a couple of helper functions in C.
But I'm pretty new to postgres internals and well I miss a reference
of functions/macro with some examples... and this is a side project
and I haven't been using C for quite a while.

Once I'll have that function I'll have to solve how to return few
rows (since I'll have to use | I expect a lot of returned rows) to
make efficient use of the gin index and avoid to compute ts_rank for
too many rows.

Don't hold your breath waiting... but let me know if you're
interested so I don't have to be the only one posting newbies
questions on pgsql-hackers ;)

[1] ts_stat could give you some hints about what lexemes may be
important... but well deciding what's important is another can of
worms... and as anticipated ts_rank should be "good enough for me".

--
Ivan Sergio Borgonovo
http://www.webthatworks.it