MongoDB Search and Sort, with Number of Matches and Exact Match

Question

I want to create a small MongoDB Search Query where I want to sort the result set based exact match followed by no. of matches.

For eg. if I have following labels

Physics
11th-Physics
JEE-IIT-Physics
Physics-Physics

Then, if I search for "Physics" it should sort as

Physics
Physics-Physics
11th-Physics
JEE-IIT-Physics

Answer 1

Looking for the sort of "scoring" you are talking about here is an excercise in "imperfect solutions". In this case, the "best fit" here starts with "text search", and "imperfect" is the term to consider first when working with the text search capabilties of MongoDB.

MongoDB is "not" a dedicated "text search" product, nor is it ( like most databases ) trying to be one. Full capabilites of "text search" is reserved for dedicated products that do that as there area of expertise. So maybe not the best fit, but "text search" is given as an option for those who can live with the limitations and don't want to implement another engine. Or Yet! At least.

With that said, let's look at what you can do with the data sample as given. First set up some data in a collection:

db.junk.insert([
    { "data": "Physics" },
    { "data": "11th-Physics" },
    { "data": "JEE-IIT-Physics" },
    { "data": "Physics-Physics" },
    { "data": "Something Unrelated" }
])

Then of course to "enable" the text search capabilties, then you need to index at least one of the fields in the document with the "text" index type:

db.junk.createIndex({ "data": "text" })

Now that is "ready to go", let's have a look at a first basic query:

db.junk.find(
   { "$text": { "$search": "\"Physics\"" } },
   { "score": { "$meta": "textScore" } }
).sort({ "score": { "$meta": "textScore" } })

That is going to give results like this:

{
    "_id" : ObjectId("55af83b964876554be823f33"),
    "data" : "Physics-Physics",
    "score" : 1.5
}
{
    "_id" : ObjectId("55af83b964876554be823f30"),
    "data" : "Physics",
    "score" : 1
}
{
    "_id" : ObjectId("55af83b964876554be823f31"),
    "data" : "11th-Physics",
    "score" : 0.75
}
{
    "_id" : ObjectId("55af83b964876554be823f32"),
    "data" : "JEE-IIT-Physics",
    "score" : 0.6666666666666666
}

So that is "close" to your desired result, but of course there is no "exact match" component. In addition, the logic here used by the text search capabilities with the $text operator means that "Physics-Physics" is the preferred match here.

This is because then engine does not recognize "non words" such as the "hyphen" in between. To it, the word "Physics" appears several times in the indexed content for the document, therefore it has a higher score.

Now the rest of your logic here depends on the application of "exact match" and what you mean by that. If you are looking for "Physics" in the string and "not" surrounded by "hyphens" or other characters then the following does not suit. But you can just match a field "value" that is "exactly" just "Physics":

db.junk.aggregate([
    { "$match": { 
        "$text": { "$search": "Physics" } 
    }},
    { "$project": {
        "data": 1,
        "score": {
           "$add": [
               { "$meta": "textScore" },
               { "$cond": [
                   { "$eq": [ "$data", "Physics" ] },
                   10,
                   0
               ]}
           ]
        }
    }},
    { "$sort": { "score": -1 } }
])

And that will give you a result that both looks at the "textScore" produced by the engine and then applies some math with a logical test. In this case where the "data" is exactly equal to "Physics" then we "weight" the score by an additional factor using $add:

{
    "_id": ObjectId("55af83b964876554be823f30"),
    "data" : "Physics",
    "score" : 11
}
{
    "_id" : ObjectId("55af83b964876554be823f33"),
    "data" : "Physics-Physics",
    "score" : 1.5
}
{
    "_id" : ObjectId("55af83b964876554be823f31"),
    "data" : "11th-Physics",
    "score" : 0.75
}
{
    "_id" : ObjectId("55af83b964876554be823f32"),
    "data" : "JEE-IIT-Physics",
    "score" : 0.6666666666666666
}

That is what the aggregation framework can do for you, by allowing manipulation of the returned data with additional conditions. The end result is passed to the $sort stage ( notice it is reversed in descending order ) to allow that new value to be to sorting key.

But the aggregation framework can really only deal with "exact matches" like this on strings. There is no facility at present to deal with regular expression matches or index positions in strings that return a meaningful value for projection. Not even a logical match. And the $regex operation is only used to "filter" in queries, so not of use here.

So if you were looking for something in a "phrase" thats was a bit more invovled than a "string equals" exact match, then the other option is using mapReduce.

This is another "imperfect" approach as the limitations of the mapReduce command mean that the "textScore" from such a query by the engine is "completely gone". While the actual documents will be selected correctly, the inherrent "ranking data" is not available to the engine. This is a by-product of how MongoDB "projects" the "score" into the document in the first place, and "projection" is not a feature available to mapReduce.

But you can "play with" the strings using JavaScript, as in my "imperfect" sample:

db.junk.mapReduce(
    function() {
        var _id = this._id,
            score = 0;

        delete this._id;

        score += this.data.indexOf(search);
        score += this.data.lastIndexOf(search);

        emit({ "score": score, "id": _id }, this);
    },
    function() {},
    { 
        "out": { "inline": 1 },
        "query": { "$text": { "$search": "Physics" } },
        "scope": { "search": "Physics" }
    }
)

Which gives results like this:

{
    "_id" : {
        "score" : 0,
        "id" : ObjectId("55af83b964876554be823f30")
    },
    "value" : {
        "data" : "Physics"
    }
},
{
    "_id" : {
        "score" : 8,
        "id" : ObjectId("55af83b964876554be823f33")
    },
    "value" : {
        "data" : "Physics-Physics"
    }
},
{
    "_id" : {
        "score" : 10,
        "id" : ObjectId("55af83b964876554be823f31")
    },
    "value" : {
        "data" : "11th-Physics"
    }
},
{
    "_id" : {
        "score" : 16,
        "id" : ObjectId("55af83b964876554be823f32")
    },
    "value" : {
        "data" : "JEE-IIT-Physics"
   }
}

My own "silly little algorithm" here is basically taking both the "first" and "last" index position of the matched string here and adding them together to produce a score. It's likely not what you really want, but the point is that if you can code your logic in JavaScript, then you can throw it at the engine to produce the desired "ranking".

The only real "trick" here to remember is that the "score" must be the "preceeding" part of the grouping "key" here, and that if including the orginal document _id value then that composite key part must be renamed, otherwise the _id will take precedence of order.

This is just part of mapReduce where as an "optimization" all output "key" values are sorted in "ascending order" before being processed by the reducer. Which of course does nothing here since we are not "aggregating", but just using the JavaScript runner and document reshaping of mapReduce in general.

---

So the overall note is, those are the available options. None of them perfect, but you might be able to live with them or even just "accept" the default engine result.

If you want more then look at external "dedicated" text search products, which would be better suited.

---

Side Note: The $text searches here are preferred over $regex because they can use an index. A "non-anchored" regular expression ( without the caret ^ ) cannot use an index optimally with MongoDB. Therefore the $text searches are generally going to be a better base for finding "words" within a phrase.