@damobrisbane could you do a pretend HTTP+JSON API example to demonstrate what you mean?

Thanks; concept only at moment. And I have to preface that I am neither a Couchdb, Machine Learning or LLM practitioner so my terminology may be bit off, but I hope not concept;

assumption is a model for a ?database exists, which has a learnt vector matrix that represents all those docs in the database.
each document includes two items: a reference to the model and a json array of floats that represents the vector for the doc.

I believe if the learnt model may be (corpus size X dimension), eg 3000X50, then the embedded vector in doc will be 50 float. Learning the model is independent of any vector embedding and adding a embedding vector to doc is also independent, but I understand if I add any new doc, the docs vector will get similarity scoring with other docs in that database, using the referenced model. I guess crucially, exactly what determines eg higher matching scores is hidden in the model, which is where all the power comes from: the model for the database, with all the learnt vector embeddings for the corpus of documents in the database, matches new documents or queries, using an inferencing process for a document on the corpus.

All this will then allow similarity matching between documents in a database. From what I can see in lucene/Couchdb this may be doable. To what end? I'm not exactly sure at the moment ;) more interested in the idea and how this type of thing might be done in Couchdb and guess from there, any possibilities that might open up.

is it distinct from an n-dimensional geo-index?