How are you handling autocomplete for short Elasticsearch queries?

Start with Elasticsearch's completion suggester before adding another service. Put useful titles and popular past queries in a dedicated suggest field, then weight suggestions by frequency or click-through.

That gives you a way to offer relevant completions while keeping retrieval in the same backend. Check how well prefix suggestions handle your actual queries before deciding you need a semantic layer.

If the problem is that people use different words from the ones in your index, a list of popular prefixes may not be enough. An intent layer could turn a fragment like a color into a structured filter before the Elasticsearch query runs.

That doesn't have to replace retrieval. But I'd test whether the extra interpretation actually improves results, and account for its latency and operating cost before putting a model in every keystroke path.

First decide what the dropdown needs to do. A fast, clear set of suggested searches might solve the request without trying to infer intent at all. The interface can be useful even if Elasticsearch is doing the ordinary search work underneath.

A command palette is another option if users need help choosing an action or narrowing a search, rather than just completing text. You can use that UI to collect parameters and still send the resulting query to Elasticsearch.