Summary
Reuters Institute reported that Nikkei built its subscriber-facing chatbot, Ask! NIKKEI, around its own archive and a custom retrieval layer because external large language models were not ready to retrieve Japanese text reliably enough on their own. The piece is a durable journalism reference because it documents a concrete newsroom architecture for archive-bounded chat rather than another generic "media company launches AI tool" announcement.
Why It Matters
This is a strong direct journalism story because it shows a newsroom making explicit operational choices about how AI should behave around reported content:
- the chatbot answers only from Nikkei's own article database
- it refuses to answer when retrieval fails instead of improvising
- it sits inside articles and suggests follow-up questions to deepen reader engagement
- the design is partly driven by copyright control and partly by language-specific retrieval limits
That makes it more useful than broad innovation coverage because it exposes the actual product and governance decisions behind a newsroom chatbot.
What the Source Says
Reuters Institute reported that Ask! NIKKEI is embedded in articles, available to subscribers, and built to answer from Nikkei's own database of articles after 2020, typically drawing on the last 18 months. Engineering manager Yosuke Suzuki said the system combines multiple relevant articles into a prompt and instructs the external LLM not to use knowledge beyond those articles. He also said the system withholds an answer when no related article can be retrieved, and that Nikkei built its own retrieval model because external LLMs were not ready to process Japanese characters well enough for the task.