Hello @sgilberg, thanks for opening an issue.

This seems like a welcome addition for the autocomplete client. I will start taking a look at this :)

Hey, I released a new version that includes the highlights metadata. Let me know if it works for you :)

Hi @Strift thank you so much for working on this! This is partly working for me, and partly not. It works pretty well for attributes that are one-to-one, and for nested arrays that don't have keys, but I'm getting errors on indexes that have nested arrays with keys (even if those arrays aren't searchable attributes).

Here's an example record to illustrate what I mean:

{
   "id": 1,
   "name": "Joseph",
   "nicknames": [
      "Joe",
      "Joey"
   ],
   "family_members": [
       {
           "relationship": "mother",
           "name": "Susan"
       },
       {
           "relationship": "father",
           "name": "John"
       }
   ]
}

In the above, the "family_members" style of data breaks the integration ("Cannot read properties of undefined (reading 'replace')" at calculateHighlightMetadata). Indexes that only have fields like id, name, or nicknames work fine.

I think there might be two parts to the issue: one is that these highlight calculations seem to be done on all fields, regardless of whether they are meant to be searchable or highlighted or not, which at best is overkill and at worst might impact performance or cause errors as in my case. The other is that, in cases where such data should be searchable, we'd want that calculation to work on those nested properties too.

For what it's worth I also notice that the match level full vs partial seems to be case dependent, and I'm not sure if that's intentional (or if it's consistent with Algolia's behavior, I'd need to dig in on that). E.g. for the example record above if I search "joseph" it will show as partial but if I search "Joseph" it is full.

Hey, I managed to get it working for the basic cases, but I want to make the logic smart enough to handle recursive scenarios. But I'm not sure if it will actually be relevant.

@meilisearch/product-team How deep in objects, arrays, or arrays of objects on average do customers have "searchable" fields? If you can share some insights with me, it would allow me to prioritize this better. Thanks 🙏