I’ve been looking into autocomplete/typeahead systems recently, especially in contexts where latency really matters (e.g. search-as-you-type or RAG pipelines).

From what I can tell, the main approaches are:

Full search backends (Elasticsearch, Meilisearch, etc.)

LLM-based suggestions (flexible but slow per keystroke)

Simpler prefix / n-gram systems (fast but sometimes limited)

I’m trying to understand what people actually use in production when you need:

very low latency

reasonable suggestion quality

minimal infra overhead

Are most systems still based on classical methods, or are people moving toward hybrid approaches (retrieval + reranking)?

For context, I’ve been experimenting with a small local implementation here:

https://github.com/MarcellM01/query-autocomplete

Not trying to replace full search systems, more to understand where the practical tradeoff line is between latency and quality.

Would be really interested to hear what setups people are running and what worked/didn’t.