qbox | NVDLA modifications for GreenSocs qbox

I'm assuming you still want to be able to click on the attachment links within the app.

So try replacing this line of code:

You says :

Search query takes around 300ms median. I am puzzled on why the server can not handle even such a low traffic.

I guess you feel that 300ms a too long. It does not mean that the server don't handle traffic, just that query is a little slow.

At first how many shard your index have The goal for a shard size should be around 50Gb. so if you have more than 1 shard and 1 replica for this 1.2GB index, it's to much.

If you can plan a maintenance on this index, you should reindex with a proper number of shards and do a forcemerge. about your mapping, not a lot of change to do.

Usually, ids are used only for strict equality, so we used to map it as keyword, not integer (user_id, id, channel_id, address_id, etc...) but it will not really have a sensitive impact on perfs

If you use a recent Elasticsearch, and don't use score, when you reindex you can think about index sorting on order_at and created_at fields. It should help saving time.
https://www.elastic.co/guide/en/elasticsearch/reference/master/index-modules-index-sorting.html

On the query itself:

You should considere to use filters. Filters don't influence score and are cachable at elastic and OS side.
In your use case, the whole index can stay in RAM. That can make a huge difference.

Your should have the same behaviour with something like this (must be tested):

I was able to reproduce your issue and reason for this is due to the use of your ngram_filter, which doesn't create any token for bob as the minimum length of the token should be 4 while standard tokenizer created bob token but then it gets filtered out in your ngram_filter where you mentioned min_gram as 4.

Even I tried with less min_gram length to 3, which would create the tokens but the issue is that both bob and bobbie will have same bob tokens, hence score for both of them will be same.

While when you search for Bobbi Smith, then bobbi ie exact token will be present only in one document, hence you get the higher score.

Note:- Please use the analyze API and explain API to inspect the tokens generated and how these are matched, this would help you to understand the issue and my explanation in details and my