Max

This feature would be incredibly useful for our work, allowing us to handle everything within Typesense without the need for additional processing. Thank you for considering it!

@kishorenc we simply concatenate the two vectors from CLIP, resulting in a 1024-dimensional embedding (512 + 512), which works pretty well for us. Although it would be great to have...

At this stage, all TS documents have 1024-dimensional embeddings. We then query for similar documents using document ID: ```json { "q": "*", "vector_query": "embedding:([], id: foobar)", "sort_by": "_vector_distance:asc, newest:desc" }

We haven’t tried the averaging approach yet since we started with concatenating and it worked well. However, I’m happy to test it within the next day to confirm. So far,...

@kishorenc We’ve tested averaging the image and text vectors, and the results are very similar. So, we’re all for using this approach in TS by default. Thank you!

@kishorenc We tested it, and CLIP embeddings are now working for multiple fields 🎉 Thank you! Could you please clarify a couple of questions: 1. Will this configuration create embeddings...

Thanks @ozanarmagan That sounds great! The only concern is whether the embeddings generation will occur again when the image value isn’t stored in the collection (`store: false`). We observe similar...