Unique vs Multiple dense vectors

Elastic Stack Elasticsearch
1

Hi there!

I'm adding semantic search to my site's search functionality. I use Elasticsearch and my index is structured as follows:

"title":{
  "type":"text",
  "analyzer":"general_analyzer",
  "fields": {
      "spell": {
         "type": "text",
         "analyzer": "spell_analyzer",
         "similarity": "BM25"
      }
  }
},
"description":{
  "type":"text",
  "analyzer":"general_analyzer"
}

I'd like to understand the best way to proceed: whether to add a `single dense_vector` (combining title + description) or to have `multiple dense_vectors` (one for each field). Why?

Note: My titles are around 60 characters / 10 words, and my descriptions are around 670 characters / 90 words.

Thank you for your attention!
[]s

2

I would start with a single dense_vector field; here’s the docs if it would help: Dense vector field type | Elasticsearch Reference

I’d probably structure the text as something like “Title: TITLE Description: DESCRIPTION” and then embed it. LLMs like having as much context as possible usually and your use-case doesn’t seem like too much context for most LLMs to me right off though your mileage will vary depending on how you do embed it.

Ultimately whether using a single field is sufficient will depend on how the data gets embedded in the space and how you want to query it. But one field is the simplest and least expensive way and will probably give you the results you want. Let me know how it goes.