Sure, I'm happy to give some insight into this.  My index itself has a few
fields - one that uniquely identifies the page, one that stores all the text
on the page, and then some others to store characteristics.  At indexing
time, the text field for each document is manually created by concatenating
each word together, separated by spaces.  Then the IndexWriter runs the
document through a custom filter that attaches payloads to each token.  The
payloads here include all the attributes I need regarding that word, and
most importantly, the index of that word on the page.  The tricky part here
was that one of my "words" could map to more than one Lucene token, so I
first create a quick map from my words to which token they should correspond
to, by running each word through an Analyzer (StandardAnalyzer in my case).
This makes it easy to only attach the payload to the first token for each of
my words.

For searching, I pass the search query to a PayloadSpanUtil which gets the
payloads for every match throughout the entire index.  I take these results
and put them into a Collection of custom objects, and then sort them first
by page identifier, and then by index on the page.  Once I have this list, I
can quickly iterate through it to find the groupings of payloads that match
the search term (this also helps weed out the occasional bad result that
comes back).  I wasn't sure initially if this would be a performance hit but
it is very quick.  Basically what I do is tokenize the search string, then
concatenate all tokens together without spaces into one string.  Then when
iterating through I see if the word matches the start of the tokenized
string - if so, chop it off and keep going til the whole string is found.
Then repeat, and so on.  It's certainly not the most elegant solution but I
didn't see a better way since PSU doesn't group or sort on its own.

One other solution I might try if I have time is to take each document from
the original search, put them one at a time into a MemoryIndex and then let
PSU act on that.  I'm not sure if this would help/hurt performance but might
be worth trying.  I will also say to make sure you apply Mark's latest patch
(see the case here: https://issues.apache.org/jira/browse/LUCENE-1465) since
it fixed some important bugs I had come across.

I hope this made sense, I haven't finished my morning coffee yet so I can't
be too sure : )  Let me know if you have any more questions.

- Greg



On Wed, Nov 26, 2008 at 3:19 AM, Eran Sevi <[EMAIL PROTECTED]> wrote:

> Hi,
> Can you please shed some light on how your final architecture looks like?
> Do you manually use the PayloadSpanUtil for each document separately?
> How did you solve the problem with phrase results?
> Thanks in advance for your time,
> Eran.
> On Tue, Nov 25, 2008 at 10:30 PM, Greg Shackles <[EMAIL PROTECTED]>
> wrote:
>
> > Just wanted to post a little follow-up here now that I've gotten through
> > implementing the system using payloads.  Execution times are phenomenal!
> > Things that took over a minute to run in my old system take fractions of
> a
> > second to run now.  I would also like to thank Mark for being very
> > responsive in fixing/patching some bugs I encountered along the way.
> >
> > - Greg
> >
>

Reply via email to