Naver to expand AI Tab with real estate, health agents as it pushes 'action-based search'

Real estate, health agents coming to AI Tab in Q3

Search-optimized model, operations tech, multimodal capabilities unveiled

Naver says AI search evolving 'from exploration to execution'

Lee Ki-chang, a director at Naver Cloud, presents at the "AI Search Tech Deep Talk" briefing. [Naver]

Naver is betting on its conversational AI search service, AI Tab, to lead the next generation of search — moving beyond summarizing results to understanding user intent and driving real-world action.

Starting this month, Naver plans to integrate real estate search and its Smart Lens visual search feature into AI Tab, with a health AI agent to follow later this year, deepening the service's role in users' daily lives. AI Tab officially launched June 25.

Naver held a press briefing Wednesday at D2SF Gangnam to outline its AI search strategy. Lee Ki-chang, a director at Naver Cloud overseeing hyperscale AI models, Han Seung-gyun, Naver's AI search service leader, and Yun Sang-du, leader of Naver's Future AI Center, attended and presented the technologies underpinning AI Tab.

Naver said it would expand AI Tab's service scope this month to deliver a next-generation search experience. "In the third quarter — starting this month — we will integrate AI Briefing and Smart Lens into AI Tab and add a real estate service," Han said. "We will also launch a Whale browser-specific agent and a health agent."

Naver also detailed the technology behind AI Tab. The service runs on what the company calls a "product-native large language model," a lightweight model built on its existing HyperCLOVA X foundation and optimized specifically for AI search. The model incorporates Naver's own data, service scenarios and user feedback into its design.

"While HyperCLOVA X is a general-purpose LLM with broad knowledge and reasoning capabilities, this model focuses on understanding conversational context, selecting the right tools for the situation, and completing the tasks users want," Lee said. "Our goal was not to top every benchmark, but to build a model that performs best at the moment a user is searching or making a purchase."

Han Seung-gyun, Naver's AI search service leader, presents at the "AI Search Tech Deep Talk" briefing. [Naver]

Naver said it rebuilt the model's training data from the ground up. "Rather than simply scraping documents from the web, we selected high-quality documents and incorporated them into training," Lee said. "We expanded the training scope beyond elementary and secondary-level knowledge to include specialized materials such as court rulings and academic papers, as well as practical everyday knowledge like product reviews, recipes and game guides."

The model's architecture was also redesigned for large-scale service. Naver adopted a mixture-of-experts structure to improve response speed and processing efficiency, and reduced end-to-end latency — the time from input to final answer. "The next-generation model maintains consistent response times even at up to 16,000 tokens," Lee said. "That means it can handle more requests with the same GPU resources, which ultimately reduces service operating costs."

Naver also refined its reinforcement learning techniques to reduce hallucinations. "We trained the model to recognize that when faced with an ambiguous question, it is better to ask for clarification than to guess," Lee said. "We also had it call actual Naver tools when generating answers. We more than doubled the computing resources allocated compared with HyperCLOVA X, and as a result the hallucination rate fell by up to 30 percentage points."

Yun Sang-du, leader of Naver's Future AI Center, presents at the "AI Search Tech Deep Talk" briefing. [Naver]

Naver also unveiled what it calls "harness engineering," the operational layer that governs how AI Tab executes tasks. "No matter how capable an LLM is, deploying it in a real service requires a separate layer of design," Han said. "Harness engineering is, simply put, the AI's operational intelligence — it instructs the LLM to retrieve up-to-date information, select the right tools, and verify its answers."

The company also outlined its multimodal search strategy, centered on Smart Lens. Multimodal technology enables AI to process and act on information in multiple formats — text, images and video — not text alone. "The visual search technology Naver has built through nearly a decade of Smart Lens is the core technology that gives our AI agents their 'eyes' to perceive the world," Yun said. "Going forward, Naver's AI agent services will evolve to understand user intent not only through text but also through images, and to connect that understanding to real-world action."

chami@heraldcorp.com