Sorting result based on dynamic terms - elasticsearch

Question

Imagine that I have an index with the following three documents representing images and their colors.
[
{
"id": 1,
"intensity": {
"red": 0.6,
"green": 0.1,
"blue": 0.3
}
},
{
"id": 2,
"intensity": {
"red": 0.5,
"green": 0.6,
"blue": 0.0
}
},
{
"id": 3,
"intensity": {
"red": 0.98,
"green": 0.0,
"blue": 0.0
}
}
]
It the user wants a "red image" (selected in a dropdown or in a “tag cloud”), it is very convenient to do a range query over the floats (maybe intensity.red > 0.5). I can also use the score of that query to get the "red-est" image ranked highest.
However, if I would like to offer free-text search, it gets harder. My solution to that would to index the documents as the following (eg use the if color > 0.5 then append(colors, color_name) at index time):
[
{
"id": 1,
"colors": ["red"]
},
{
"id": 2,
"colors": ["green", "red"]
}
{
"id": 3,
"colors": ["red"]
}
]
I could now use a query_string or a match on the colors field and then search for "red", but all of a sudden I lost my ranking possibilities. ID 3 is far more red than ID 1 (0.98 vs 0.6) but the score will be similar?
My question is: Can I have the cake and eat it too?
One solution I see is to have one index that turns free text into "keywords" which I later use in the actual search.
POST image_tag_index/_search {query: "redish"} -> [ "red" ]
POST images/_search {query: {"red" > 0.5}} -> [ {id: 1}, {id: 3}]
But then I need to fire two searches for every search, but maybe that is the only option?

Answer 1

You can make use of nested data type along with function_score query to get the desired result.
You need to change the way you are storing image data. The mapping will be as below:
PUT test
{
"mappings": {
"_doc": {
"properties": {
"id": {
"type": "integer"
},
"image": {
"type": "nested",
"properties": {
"color": {
"type": "text",
"fields": {
"keyword": {
"type": "keyword"
}
}
},
"intensity": {
"type": "float"
}
}
}
}
}
}
}
Index the image data as below:
PUT test/_doc/1
{
"id": 1,
"image": [
{
"color": "red",
"intensity": 0.6
},
{
"color": "green",
"intensity": 0.1
},
{
"color": "blue",
"intensity": 0.3
}
]
}
The above corresponds to the first image data the you posted in the question. Similarly you can index other images data.
Now when a user search for red the query should be build as below:
{
"query": {
"bool": {
"must": [
{
"nested": {
"path": "image",
"query": {
"function_score": {
"query": {
"bool": {
"must": [
{
"match": {
"image.color": "red"
}
},
{
"range": {
"image.intensity": {
"gt": 0.5
}
}
}
]
}
},
"field_value_factor": {
"field": "image.intensity",
"modifier": "none",
"missing": 0
}
}
}
}
}
]
}
}
}
You can see in the above query that I have used the field value of image.intensity to calculate the score.