Localised by Design: How GoTo Built Upon SEA-LION Embeddings to Train A Model That Understands Indonesia

ByAISG | AI Products
Fluent Isn’t the Same as Understood

How GoTo taught AI to think in Indonesian, not just speak it, building upon AI Singapore’s SEA-LION embedding models

AI Singapore × GoTo Group

Basic language capability can handle grammar just fine, but true linguistic fluency is what gets the meaning right, and not all AI models can make that leap. To be truly helpful, an AI model needs to understand you, including your lingo, your cultural references and your unique way of speaking. Otherwise, it will still misunderstand user intent, recommend the wrong restaurant or send a delivery to the wrong address.

Models launch as “multilingual,” and by the narrow test of grammar, they are. But understanding a language and understanding a place are different achievements, and the gap between them shows up exactly where it costs the most: assuming your intent before any output is generated. 

GoTo Group, the largest digital ecosystem in Indonesia, comprising Gojek and GoPay, and serving hundreds of millions of people across Indonesia, endeavors to bridge this gap through their AI-powered platforms. Instead of a quick fix, GoTo’s teams rebuilt the layer underneath: the embedding model that determines what a query means before the rest of the system responds. Working with AI Singapore’s open-source SEA-LION embedding models, GoTo’s scientists built a production-level embedding model that matches or beats strong global AI models on its own home turf by natively grasping Indonesian context that global models frequently miss.

A Word That Means Two Different Things at Once

Take the word “gacor”. To most Indonesians, it’s slang for “amazing” representing an emotion of being in awe of hitting a jackpot, someone pulling off a great outfit, or a stand-out performance. On Gojek’s (ride-hailing) platform, it means something else entirely: a driver-partner whose app is flooded with ride and delivery orders. Or take another word: “anyep”, everyday Indonesian for bland, grey, or lifeless. Inside the platform, it flags a driver-partner who hasn’t received a single order all day.

These aren’t isolated linguistic exceptions; they are the operational vocabulary a business runs on. A model trained only on the open web only ever learns the public meaning, because the alternate meaning doesn’t exist on the web, yet exists within the product.

The Layer Nobody Talks About

Most conversations about AI in SEA focus on the visible layer: whether a chatbot replies in the user’s local language. Far fewer talk about the layer sitting underneath almost every AI-powered search, recommendation, or retrieval-augmented generation (RAG) system: the embedding model.

Embedding models don’t generate language. Instead, they convert text such as a query, a menu item, or a half-typed address into numbers representing its meaning. Similar meanings land close together in the semantic latent space, unrelated ones land far apart, and search works by finding the nearest neighbours. An embedding model is like a librarian who shelves books not alphabetically, but by meaning, so that related ideas end up physically close together. Except here, an embedding model shelves words, images, and concepts instead of books.

No amount of prompting fixes a retrieval problem; the damage is done one layer down. If you want an AI product to understand SEA rather than just speak it, the embedding model is where that understanding can be built into the system, preventing problems from arising in the first place.

Why the Gap Is Structural, Not a Detail to Patch

This is not a gap that more training data can simply close. Firstly, SEA’s diversity is difficult to represent–with 11 countries, over 700 million people, and more than 1,200 languages, compounded by constant code-switching such as Indoglish and Singlish, that shifts meaning mid-sentence. Secondly, 90% of the global training data is skewed to English or a handful of dominant languages, leaving SEA languages barely visible.

To many, the obvious fix is to take a strong global model and train it further on local data. But adapting a global model is harder than it looks, and when that local data is very different from what the model originally learned, three things tend to go wrong:

  • The model forgets. Training hard on unfamiliar data can erode knowledge it already had,  leading to catastrophic forgetting.
  • Its “map of meaning” is the wrong shape. A global AI model naturally links the word “driver” to “car”. But in a ride-hailing app, “driver” also needs to connect to “customer” and “tips.” Rearranging that map after the fact is unreliable and expensive.
  • Teaching models from scratch is costly. Instilling regional knowledge through hand-labelled data requires heavy annotation spend, which is far more than having the regional knowledge introduced during the model’s foundational training phase.

GoTo’s decision: do not retrofit representation. Start from a foundation that is already regionally native, and specialise from there.

Starting From a Southeast Asian Foundation

That foundation was AI Singapore’s open SEA-LION ModernBERT models. These compact encoders (300M and 600M parameters) with Gemma 3’s tokeniser, pretrained on 245 million text pairs across 11 regional languages, including Burmese, Chinese, Filipino, Indonesian, Khmer, Lao, Malay, Tamil, Thai and Vietnamese. Three things made it the right starting point:

  • Already regional as a base. The embedding model starts with working knowledge of SEA languages and scripts, so further training reshapes an already-relevant map instead of building one from nothing.
  • Small and efficient. At 300M and 600M parameters, it runs on modest hardware and can be deployed on a company’s own infrastructure, providing a real advantage for cost, latency, and data residency.
  • Open and adaptable. As an open-source model, SEA-LION can be freely retrained on private data, a flexibility closed models do not offer.

Four Stages, One Direction

GoTo adapted the model in four progressive stages, moving from general language competence to sharp, task-specific search:

  1. Relearn the language by filling in blanked-out words to refresh command of the relevant languages.
  2. Focus on quality through the same training, now on cleaner, higher-quality text.
  3. Learn what counts as similar, pulling matching meanings together and pushing unrelated ones apart.
  4. Tune for real tasks, resulting in a final stage optimised for domain- and language-specific tasks.

The mix leaned mostly on general Indonesian and English text, plus a small, deliberately curated share of GoTo’s own in-domain data, built from unidentifiable, yet genuine user behaviour. It matches what someone searched with their intended selection rather than labels alone, capturing the messy reality of abbreviations and informal spellings.

Where catalogue descriptions were thin, GoTo used GoTo-AI embedding model , an Indonesian-specialised model built on SEA-LION, to generate short descriptions, and close the loop within the regional open-source ecosystem. The turnaround required just a few hours on a modest multi-GPU setup, not weeks of retraining.

The Results: Strongest Improvement Where Local Interpretation is Critical

GoTo-AI  embedding models matched or outperformed leading global models across its own search tasks, and the size of the gap tracked almost perfectly with how “local” each task was.

Search TaskBest Global ModelGoTo’s Embedding Model
Location searchStrongClearly ahead ↑
Restaurant searchStrongAhead ↑
Payment searchModerateFar ahead ↑
Product searchStrongAhead ↑
General help articlesStrongRoughly level ≈

On tasks dense with Indonesian names, abbreviations, and platform-specific slang such as bills, places and products, the adapted model pulled decisively ahead. On the most generic task, general help articles, the global model held a narrow edge. This serves as an honest signal that local adaptation pays off most where local knowledge is the deciding factor.

The Takeaway For Building With Localisation

The layer that critically decides whether an AI product understands its users, namely retrieval and embeddings, is invisible, easy to leave out on a foreign-trained default, and the first-place representation quietly breaks down to the detriment of downstream tasks that truly matter. For SMEs and enterprise teams building search, recommendation, and RAG systems across Indonesia, Vietnam, Thailand, the Philippines, and beyond, the lesson is not to “use a bigger model.” Instead, it is a design choice made earlier than most teams think to make it:

Choose a foundation — model, tokeniser, and embedding space — that already represents the region, and let every layer built on top of it inherit that advantage.

Retrofitting representation after launch is always more expensive and less complete than starting from a base that never had to unlearn a foreign-first view of the world. That’s the shift AI Singapore built SEA-LION’s embedding suite to enable. Unlike a full LLM, embedding models cost a fraction of the amount to train, and because SEA-LION’s embedding models are independent models, builders across the region can swap them in without rebuilding their stack. The gains are immediate, and the intent is clear: not as a fix applied after the fact, but as a starting assumption from day one.

Build With SEA-LION

The SEA-LION embedding models are open and free for research and commercial use.

If you’re building AI for Southeast Asia and want a foundation that already understands the region, we’d love to hear from you.

With thanks to the GoTo-AI team and the AI Singapore embedding team for the collaboration.