Vectors caching

[bpolaszek]

Consider the following document:

{
  "id": 1,
  "productId": "A"
  "name": "Unicorn T-shirt"
  "color": "Red"
  "size": "L"
}

Consider I have an embedder with documentTemplate set as {{doc.color}} {{doc.name}}. It will generate embeddings for Red Unicorn T-shirt.

Now, if I add the following document:

{
  "id": 2,
  "productId": "A"
  "name": "Unicorn T-shirt"
  "color": "Red"
  "size": "S"
}

In this case, Meilisearch would also call the embedder to get vectors for Red Unicorn T-shirt.
Which would be the same as the 1st product variant.

On huge catalogs with lots of variants, tons of embedder calls could be avoided if some cache mechanism were implemented. The cache key would be sth like sha1(resolve(documentTemplate))-sha1(embedderSettings) (unsure if the index name should be part of it). That cache key would be stored along with each document (meaning, the last document holding the hash being deleted would also delete the cache entry, for example).

WDYT?

[curquiza]

Hello @bpolaszek
Thanks for the request
For your information, feature requests are handle here. I've moved it to the right place, Thank youuu