One of the things that was still on my wish list for this site, was a proper search. In two articles I will explain how I’ve done this. The next article will describe the way I have currently set things up. This article will primarily focus on the journey I made to come to my choice (Djapian).
Coming from Plone, I’m quite spoiled in the search department: Plone has great search functionality out of the box. When launching the Django version of this site, I didn’t want to spend too much time on the search feature since it isn’t used that much, according to the access logs.
As a result I implemented the simple solution where I just searched for weblog entries containing the search term:
BlogEntry.objects.published().filter( Q(title__icontains=query) | Q(intro__icontains=query) | Q(body__icontains=query))
This works for simple queries, but it’s far from perfect since it literally searches for a string. Searching for “plone django” for example only returns articles where I used the word “plone” and then “django” separated by a space.
In my quest to improve the search function, I started looking for Django applications to help me out. Although I didn’t fully research this area, Haystack seemed to be the right tool for the job.
And it’s indeed a really nice application to use. A few highlights:
- The clear documentation gets you up-and-running pretty quickly.
- Haystack is using a class based views, so it’s really easy to customize.
- You can use a data template to feed the indexer the right data. (Which implies that you can use template tags to e.g. strip out HTML tags. And since I’m storing the body text of my entries as HTML, this can be useful.)
Haystack provides a nice API to search but you still have to decide on
which backend to use. Initially I chose Whoosh
since it’s pure Python. (This makes is easier to integrate in my
buildout configuration.) I used
as that is the latest version on PyPI. However, using the last release
of Whoosh results in an
AttributeError when reindexing:
AttributeError: 'FileIndex' object has no attribute 'commit'
This means I had to use Whoosh 0.3.18 since Haystack 1.0.1 does not seem to be compatible with the Whoosh 1.x series.
And to be fair: searching indeed worked. But when I viewed the results, the score was always zero. Which isn’t useful if you want to sort on that score. (Note that this might be due to the fact that I’m not using the most recent code of both Whoosh and Haystack. I did not want to spend time on researching this to be honest.)
I figured that switching backends might be an easy solution. So I chose Xapian in combination with xapian-haystack version 1.1.3beta. (Mainly because I didn’t like the idea of running Solr.) And initially everything seemed to work the way I wanted. That is: searching for “plone django” gave me results and those results had a nice score I could sort on.
But there was no difference between the results of “plone AND djano” and “plone OR django”. And worse: searching for “postgres” or “postgres*” did not return articles containing the string “PostgreSQL”. This was not acceptable.
I know that the Xapian backend supports all kinds of feature flags to enable these kind of things, but I could not immediately figure out how to do it via Haystack.
An article by Nizam Sayeed on his blog Nomad Journey about full text search in Django (original link is missing) sent me in the direction of Djapian. And judging by the subversion logs of the code, it’s still actively developed.
This switch from Haystack to Djapian meant that I could not use the class bases view. Then again, I did not have to either. Using Djapian meant that I only had to make relatively small changes in my original code. But that’s something for a next article.
By the way: the documentation of Djapian may look less polished than Haystack’s, but it contained all the information I needed.
There’s more than one way to implement search in a Django site. The Djapian/Xapian based approach is a good solution for me. Perhaps I could have ended up with a similar result with Haystack/Whoosh if I had wanted to. This just seemed liked the easiest way to go for now.