Hi all, I’m currently working on using the

embed

feature in Typesense 0.25 for semantic search. Here’s a part of my schema: { "name" : "embedding_multi", "type" : "float[]", "embed": { "from": [ "description", "item_details_en" ], "model_config": { "model_name": "ts/multilingual-e5-base" } } }, In this setup: •

description

is a regular string (e.g., product description) •

item_details_en

is an array of objects like: [ { "name": "Material", "title": "Stainless Steel" }, { "name": "Color", "title": "Black" } ] I would like to ask: • How does Typesense internally convert these values into a single string before sending them to the embedding model? • Will the values in the object array be concatenated, and in what format (e.g.,

"Material: Stainless Steel\nColor: Black"

or just raw values)? • Do I need to pre-process

item_details_en

into a string manually, or is it handled internally?

I'd suggest using dot notation to specify the sub-field itself :

item_details_en.title

The combine all field for the embedding using "space" or what?

Typesense takes all specified fields in the

from

array, then concatenates their contents with a space separator to form a single text string and finally sends this combined text to the embedding model for processing For example, if you have:

"embedding": {
  "type": "float[]",
  "embed": {
    "from": ["product_description", "product_name"],
    "model_config": {"model_name": "ts/e5-small"}
  }
}

The values from both

product_description

and

product_name

would be joined with a space between them, and this combined text would generate a single embedding vector. It handles both string and string array types, and properly skips missing fields.