lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Uwe Schindler" <>
Subject RE: best practice: 1.4 billions documents
Date Fri, 26 Nov 2010 17:49:46 GMT
This is the problem for Fuzzy: each searcher expands the fuzzy query to a
different Boolean Query and so the scores are not comparable - MultiSearcher
(but not Solr) tries to combine the resulting rewritten queries into one
query, so every searcher has the same query.

And here starts the second bug: If one of the clauses of a BQ are negative
MTQs (MUST_NOT), then the result is wrong - and this is not fixable as it
not only affects BQ. Also MTQs that rewrite to Span (like the new
SpanOrWrapper) are totally wrong combined. The problem is that the negative
clauses are only correct for the searcher they were created for. If you pass
the rewritten query to another searcher (what MultiSearcher does) it may not
exclude all needed documents, because some TermQueries or whatever are
missing (because the got no term hit in the first searcher). The combine
method is too stupid and does not handle that correctly.

And the latter is the big problem: (sorry the issue is not
filled with all necessary information, as Robert an Me had lots of sleepless
nights and discussions and tests done and came to the final answer: It's
unfixable :( for all queries)

Uwe Schindler
H.-H.-Meier-Allee 63, D-28213 Bremen

> -----Original Message-----
> From: [] On Behalf Of Yonik
> Seeley
> Sent: Friday, November 26, 2010 3:28 PM
> To:; Uwe Schindler
> Subject: Re: best practice: 1.4 billions documents
> On Mon, Nov 22, 2010 at 12:49 PM, Uwe Schindler <> wrote:
> >  (Fuzzy scores on
> > MultiSearcher and Solr are totally wrong because each shard uses
> > another rewritten query).
> Hmmm, really?  I thought that fuzzy scoring should just rely on edit
> Oh wait, I think I see - it's because we can use a hard cutoff for the
number of
> expansions rather than an edit distance cutoff.  If we used the latter,
> should be fine?
> The fuzzy issue I would classify as "working as designed".  Either that,
or classify
> FuzzyQuery as broken.  A cuttoff based on number of terms will yield
> results even on a single index.  Consider this scenario: it's possible to
add more
> docs to a single index and have the same fuzzy query return fewer docs
than it
> did before!
> -Yonik

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message