Elasticsearch works well once it receives a complete query, but many users only type one or two words. That makes it feel like the bigger challenge is helping people express what they want before the search runs. I'm exploring ways to add autocomplete or an intent layer in the UI while keeping Elasticsearch as the retrieval backend, rather than replacing it with a hosted service whose pricing can grow quickly.
Some lightweight search engines look useful, but they seem more like backend replacements than solutions to the vocabulary and intent problem. Has anyone built this kind of flow? Did you mainly improve Elasticsearch queries and suggesters on the backend, or add a frontend layer that turns partial input into filters, parameters, or expanded queries?
3 Answers
Before adding an AI service, consider Elasticsearch’s completion suggester. You can index popular past searches, titles, or other phrases in a dedicated suggest field and rank them by frequency or click-through rate. It’s fast and handles the common prefix-autocomplete case well. An embedding or LLM layer is probably better reserved for genuinely semantic intent that typo-tolerant matching cannot cover.
A hybrid approach can work: use a lightweight intent step to turn fragments into structured filters and possible synonyms, then send those to Elasticsearch. The final query could combine boosted lexical matches, the extracted filters, and semantic or vector matching at a lower weight. That keeps Elasticsearch as the source of retrieval while improving results for vague two-word searches, though you’ll need to watch latency, caching, and inference costs.
Don’t overlook the UX side. Many users may simply want a fast, useful dropdown with suggested queries, categories, and filters rather than a full natural-language interpretation system. A responsive autocomplete component backed by cached popular searches can deliver a lot of value before you introduce more complicated backend or AI infrastructure.
The best balance is probably to keep the input simple and fast while gradually adding richer suggestions based on what people actually search for. That gives you real usage data before committing to a larger intent architecture.
I’d start with deterministic suggestions and filters, then add an AI layer only for cases where the user’s intent is genuinely ambiguous. Calling a model on every keystroke could make the interface feel slower and become expensive quickly.