Thinking Sphinx And Unicode
For a major project at my current day job, we’re using the excellent Thinking Sphinx Ruby on Rails plugin to provide speedy and useful search functionality.
The Thinking Sphinx developer has an introduction over on his blog but here’s the condensed version. Sphinx is a separate daemon that runs on your system which indexes a range of attributes and fields back to an integer ‘document id’. In Rails terms, this means it can index 1 or more of your model attributes to the model id.
What’s the point when you can just do an find() call on the model and let the database do the searching? Sphinx is blindingly fast (even with very large datasets), it can sort by relevance, and it can sort across models.
There are three characteristic of our current project that recently made me delve into some of the complexities of sphinx.
- Our main database was carefully constructed to be utf-8 encoded
- The database primarily contains data on books. Generally English language, but plenty of the titles and authors have non English characters in them
- Our office is in Australia, so all our users have English language keyboards
Say we have an author whose name is “Alex Čihař”. Authors keep us in business, so we’d prefer to keep them happy. Therefore, mangling their name on our promotional material is probably something to avoid, so we store Alex’s name in our database accents and all and give ourselves a pat on the back.
Then one of our clients calls customer service and asks for a copy of the new book from Mr Čihař. Our staff member does their best with the keyboard at hand and searches the system for “Alex Cihar”.
I bet you can see where this is going. No results.
The trick here is to use a thing called character folding. The Unicode consortium have a technical report on the topic if you feel like some bedtime reading. The long and the short of it is that in our particular environment, we wanted all of the following characters to be equivalent when searching: C c Ć ć Ĉ ĉ Ċ ċ Č č, and much the same for the accented versions of all the other latin-ish looking characters.
Lucky for us, Sphinx has a feature called charset_table for just this situation. It’s purpose in life is to map one Unicode character to another and make them functionally equivalent in the indexes.
With some processing of a CSV file provided by the Unicode consortium, it was possible to build a reasonable value for the charset_table option.
In Thinking Sphinx, it should go into your config/sphinx.yml file (with no line breaks):
This charset_table option is saying that:
- The characters 0->9, a->z and _ should be considered their usual values
- A should be considered equal to a
- M should be considered equal to m
- Unicode Codepoint U+00C0 (À) should be considered equal to a.
- Unicode Codepoint U+1EEC (Ử) should be considered equal to u.
Anything not listed in the table (ie. ß, punctuation, Hebrew characters, Japanese characters) is considered white space and cannot be searched for. If these extra characters are important to you, don’t forget to add them in.
If you’re not using Thinking Sphinx, then the value can go directly into your Sphinx config file under the appropriate index. Without Thinking Sphinx, you should also make sure that your index is set to utf-8 mode using the charset_type setting.
Mr Čihař is happy, our customer is happy, and we don’t lose a potential sale. Winners all around.
Update, 23rd September 2008:
- Thanks to an email from David Krmpotic, I’ve added a few missing entries to the mapping list.
Update, 18th June 2009:
- Thanks to an email on the Thinking Sphinx list from Marcin, I’ve added a the polish character ł
Update, 8th March 2011:
- Thanks to an email from Clement Hallet, I’ve added ø and Ø as synonyms for o.
Update, 14th December 2013:
- Thanks to an email from Paweł Bator, I’ve added the polish character Ł.
Update, 1st March 2014:
- Thanks to an email from Marko Ojleski, I’ve added Đ and đ as synonyms for d.