Bilingual AI Tour Guide: A LINE Official Account for Multilingual Visitor Support

When a foreigner walks into a Taiwanese motor vehicle office

An English-speaking business traveler walks into a motor vehicle office, passport and international driver’s license in hand, hoping to convert it into a local license.

What greets him: signage entirely in Chinese, an unfamiliar ticket-number system, unclear service counters, and front-desk staff with limited English. He might end up waiting twenty minutes in the wrong line, only to discover he’s missing a document.

This isn’t an isolated case. As Taiwan advances its bilingual-nation policy, genuinely serving foreign residents has become a real, practical challenge for government agencies.

---

Project background

This project grew out of a government initiative to promote a “cross-agency bilingual-friendly environment.” A city/county motor vehicle office wanted to offer foreign residents instant, accurate English-language service — without adding to front-desk staff workload.

Two core requirements stood out:

1. Let residents ask in English and get an answer correctly mapped to Chinese-language regulations
2. Ensure every AI answer is grounded in an official source — never wrong information

The second point is the single most critical constraint for AI applications in a government context. A wrong regulatory answer could send a resident away for nothing — or even delay their case.

---

Why a LINE Official Account

Given how foreign residents actually use their phones, the entry point needed no new app install:

QR code posted on-site → scan to add the LINE Official Account → type a question → get an instant bilingual answer

The entire interaction happens inside LINE, so residents never have to learn a new interface — the friction is close to zero.

---

The technical core

A typical AI chatbot has a “hallucination” problem — the model can confidently give a wrong answer when it’s uncertain. In a regulatory-guidance context, that’s an unacceptable risk.

This project’s design uses a RAG (Retrieval-Augmented Generation) architecture, with the following flow:

1. The resident asks a question in English
2. The AI parses the intent in real time (semantic understanding, not keyword matching)
3. It retrieves the precisely relevant regulation passages from the approved Traditional Chinese knowledge base
4. It generates an English answer grounded in that retrieved source text
5. The answer includes the source document and citation, so the resident can verify it themselves

This “retrieve first, then generate” mechanism ensures every sentence the AI produces is backed by an official document — not a guess pulled from the model’s memory.

---

Five core modules

1. Knowledge Asset Management (KAM)

Centrally manages all service knowledge — license conversion procedures, required documents, fee schedules, processing times, and more. Supports version control: once an administrator updates a document, it takes effect immediately, so the AI always works from the latest approved version and never falls back on outdated information.

2. Retrieval-Augmented Generation Engine (RAG Engine)

Uses semantic understanding rather than keyword matching, capturing what residents actually mean. Even when a resident phrases a question in casual English, the system can still map it to the correct Traditional Chinese regulation passage.

3. Answer Quality Governance (AQG)

Staff can rate and correct every AI answer. Manual corrections feed back into optimizing the system prompt — no model retraining required, balancing efficiency with data security. This lets the AI’s accuracy keep improving over time.

4. Query Behavior & Trend Analytics (Analytics)

The system logs every query and analyzes which questions come up most often — and which ones the AI can’t answer (knowledge gaps). This lets the agency proactively identify where the knowledge base needs expansion, instead of waiting for complaints.

5. Platform Governance & Auditing (Governance)

Every query, retrieval, and answer is fully logged, meeting government agencies’ compliance and information-security requirements.

---

A few key discussions

The cross-language challenge: Chinese knowledge base, English questions

All of the motor vehicle office’s official documents are in Traditional Chinese, but residents ask in English. The RAG system needs to bridge that language gap through semantic understanding rather than literal translation, finding the right matching content in the Chinese knowledge base. This was one of the core considerations in this project’s technical design.

The AI can’t just say “I’m not sure” and leave it there

In a general conversational context, an AI saying “I’m not sure, please contact the relevant department” is a reasonable answer. But at a motor vehicle office counter, the resident is standing there waiting — that kind of answer helps no one. So the design requires that when the system can’t retrieve supporting content from the knowledge base, it explicitly tells the resident, “This question is outside what the system can answer — please ask at the service counter,” and points them to the relevant counter number.

The labor cost of maintaining the knowledge base

Regulation changes and form updates are routine. If a technical staff member had to be involved every time the knowledge base needed updating, the long-term maintenance cost would deter the agency from adopting the system at all. This project was designed so service staff can upload and update documents themselves through the back office, keeping the technical barrier as low as possible.

---

System architecture

• Front-end entry point: LINE Official Account (LIFF or Webhook)
• AI engine: RAG architecture, LLM + vector retrieval
• Knowledge base: official Traditional Chinese documents (procedure guides, regulations, form instructions)
• Cross-language processing: English semantic understanding → Traditional Chinese knowledge base retrieval → English answer generation
• Back office: knowledge management interface (usable by non-technical staff), audit logs, answer rating

---

The core lesson from this project

When a government agency adopts AI, “how accurate is the answer” matters more than “how smart is the answer.”

Residents don’t need an AI that can chat — they need an AI that speaks with evidence and never leads them astray. That’s exactly where the value of a RAG architecture lies — it’s not about making the model smarter, it’s about making every single answer verifiable against a source.

This project made something clearer to us: in any AI application touching regulations, procedures, or official information, “knowledge governance” matters just as much as the model’s own capability.