Using 'match' instead of 'term' in search queries

[Treora]

After I just had an issue with a capitalised uri not returning any results from an annotation search (using a non-default mapping), @tilgovi suggested to use 'match' instead of 'term' filters in queries. Would this be an improvement, and would it provide a solution to #73 and #75 as well?

As explained on stackoverflow:

The match query will apply the same standard analyzer to the search term and will therefore match what is stored in the index. The term query does not apply any analyzers to the search term, so will only look for that exact term in the index.

A related question is: What is the desired behaviour of the search? For a URI, userid and such an exact match seems appropriate, while perhaps for the text field fuzzy matching could be acceptable? I suppose that these things could be controlled in the mapping.

For reference, right now the main part of queries is built like this: (in elasticsearch.py)

    for k, v in query.iteritems():
        q['filtered']['filter']['and'].append({'term': {k: v}})

[tilgovi]

cc @birdsarah, @nickstenning

I think using match queries instead of term queries makes a ton of sense. The only instance where I think we would want a term query the mapping would already specify not_analyzed for the field, which makes the match and term queries equivalent.

Whether to use the fuzzy matching support in Elasticsearch is another question entirely. I'd rather not, for now. Applications can always subclass the model and build queries in a different way if they really need to, or use the raw search API.

[Treora]

Ok I finally looked into Elasticsearch a bit more. I think we should just use term and make all fields not_analyzed by default (or equivalently using the keyword analyzer as the default). This makes that searches will find only exact matches (case-sensitive), which seems like a good default behaviour without a risk of surprises.

Just not sure which fields should be analyzed. Most have not_analyzed specified, which can be cleaned out when making this the default. Should the others be analysed with the standard analyser?

[nickstenning]

Closed by #88.