IsonSearch, our search engine for the Indonesian web, now retrieves pages by meaning as well as by keywords. A question phrased in everyday Indonesian can land on an authoritative page written in formal administrative language, even when the two share almost no words.
This note explains how it works and what it changes for IsonAI, the assistant that uses the index as a grounding source for its answers.
Why keywords are not enough
Classic search matches words, not meaning. It works well when the query shares vocabulary with the page it should find. For Indonesian, that assumption fails more often than it should.
People ask in an everyday, colloquial register, while authoritative pages are written in formal administrative language. Someone types "cara ngurus KTP hilang". The government page that actually answers the question says something like "prosedur penerbitan KTP-el karena hilang". Same intent, almost no shared words, so a keyword engine can miss the best source entirely.
The register gap is not the only one. Indonesian text mixes synonyms and loanwords side by side (daring next to online, unggah next to upload), leans on abbreviations everywhere (BPJS, SIM, NPWP, pemda), varies in wording from region to region, and is often typed with casual spelling.
For a person this is a small annoyance. You rephrase and search again. An AI assistant retrieving evidence in one shot does not get that second chance, so a missed source becomes a weaker answer.
Retrieval by meaning, next to keywords
Pages in the IsonSearch index now also receive a vector representation, an embedding that captures what the page is about. The embeddings are produced by an AI model we run on our own hardware.
A query is embedded the same way. The engine then looks for the pages whose vectors sit closest to the query vector, a k-nearest-neighbor (KNN) search. Distance in that space follows meaning rather than spelling, which is exactly what the examples above need. A colloquial question and a formally written page end up near each other because they are about the same thing.
Meaning-based results do not replace keyword results. The two are combined. Lexical matching stays precise where it is strong, on names, numbers, and exact phrases. The semantic side adds recall where keywords fail, on paraphrase and register gaps.
Curation still applies on top of both. Source-quality tiers place government, established media, and academic sources above the general web, and the filtering done by IsonSearch Curator decides what can rank at all. Semantic search widens what we can find. Curation keeps what we return trustworthy. We describe that layer in Curating the Indonesian web for AI.
Making it fit on hardware we own
The index covers millions of curated Indonesian pages. Storing a vector for every one of them, and searching those vectors quickly, is the hard part of this work, because we do it on our own machines in Jakarta. Renting a vector database in someone else's cloud was not something we were willing to do. The index and the queries stay in-country.
The design that made it fit has two parts. First, vectors are stored in a compact form. Second, search runs in two stages: a fast first pass over the compact vectors builds a shortlist, then a precise second pass re-scores that shortlist. The shortcut only helps if it does not change what users see, so before serving it we checked the two-stage search against exact search on real queries and confirmed that in practice the results hold up.
Freshness is handled at ingest. New pages receive their vectors as they enter the index, so meaning-based search stays as fresh as the crawl itself.
What changes for IsonAI
IsonSearch is a grounding source for IsonAI. When someone asks the assistant about an Indonesian topic, it retrieves evidence from the index and cites it.
With semantic retrieval live, a question asked in everyday words can now land on the authoritative page phrased formally. In practice that means fewer missed local sources and better-grounded answers, with citations that lead back to real Indonesian pages.
Semantic retrieval is one leg of ranking, not the whole of it. It works together with the rest of retrieval, including freshness signals and the quality tiers described above. It is running in production today, serving both IsonSearch and IsonAI.
Where it stands
Meaning-based retrieval is live and improving. The embeddings, the blend between lexical and semantic ranking, and the evaluation behind them continue to be tuned.
IsonSearch itself is introduced in Introducing IsonSearch, and a shorter companion note on this feature is at IsonSearch now searches by meaning. Questions and research inquiries are welcome at research@epithre.com.