Do LLMs replace the need for Elasticsearch?

It depends on the use-case :-) but I’d say no. The use cases I’ve seen are either people doing prompt engineering to create queries in which case you’re using es the standard way and have a query understanding layer

Or people do some fine tuning of a model or their own and store it someplace (maybe es). In which case you could still be using es to cover filters and range parts of the queries and then rerank that result set.

Not covered the ops and scalability benefits you get with elastic / solr / open search which is also very important if you’ve non trivial usage

https://www.elastic.co/what-is/vector-search