A chatbot that answers only from verified sources
Affordable, secure, fully customizable
It answers only from verified material to minimize hallucinations, and we tailor the tone, mood, answer length and greetings exactly the way your organization wants.
No external AI, zero cost per question
It never calls an external AI, so there are no token costs, and it runs without a GPU.
Your data stays inside
Questions and documents never leave your server, and question logs keep no IP addresses or account details.
Customize it your way
Formal or friendly speech, emoji, answer length and forms of address are all set in a single config file. You can even preview other tones right in the chat screen.
Updates without downtime
Edit FAQs, answers or tone in the admin screen and deploy — the chatbot answers with the new content from the very next question, without a restart. Changes are automatically checked before deployment and can be rolled back in one step. We also make every effort to apply engine and security updates without downtime; if a short maintenance window is ever needed, we announce it in advance.
Try demos by field
Pick a field — district office, foreign residents, university and more — and start asking. You can also see the widget that goes on your website.
Choose a demo bot →Pricing
License, extra languages, tuning and maintenance prices, minimum specs and the rollout process — all on one page.
See pricing →Request a demo · Contact
Ask for a demo built from your materials, a quote, or installation advice.
Contact us →Exactly what an institutional chatbot needs — done right
Generative AI chatbots sound convincing, but for public and institutional guidance, never being wrong and never letting data leave the institution come first. narubot™ was built on those two principles.
Proprietary technology
An in-house answer engine that doesn't borrow external AI
narubot™'s answer engine is not a service layered on top of third-party generative AI (ChatGPT, Claude, Gemini, etc.). From understanding the question to finding the evidence and composing the answer, an engine we designed and developed ourselves produces every answer inside the institution's server. We call this approach VGA — Verified Grounded Answering.
Zero external dependencies
No external AI models, APIs or agent services. Vendor price hikes, policy changes and service shutdowns don't affect you.
We own the engine
We manage the engine code ourselves, and versions and file checksums let us confirm exactly which engine is installed at which institution.
Quality by design, not headcount
Automated validation, evaluation questions, version history and zero-downtime deployment protect quality. It runs reliably without a large team.
Built for your language
Tuned so that different wording, spacing, typos or follow-up questions still reach the same answer.
Security
Not a single character of questions or documents leaves the institution
Third-party generative AI chatbots send questions to external servers to get answers. narubot™ produces every answer inside the institution's server, so citizens' questions and internal documents are never sent to an external AI company — or to us.
🏢 On-premises · closed networks
Installed directly on the institution's server and works without an internet connection. An installation bundle for closed networks is provided.
🚫 Zero external transmission
With no calls to external AI or agent APIs, there is simply no path for questions or documents to leave the institution.
🙈 Minimal personal data
Question logs store no IP addresses or account information. Inquiry contents are automatically deleted after one year.
🔐 Admin protection
Passwords are stored only as hashes, and every input form has CSRF protection, login-attempt limits and session expiry.
✅ Validate, deploy · roll back
Edited answers go live only after passing automated checks, and if there is a problem you can revert to the previous state in one step. Every change is logged.
📎 Upload checks
Uploaded documents are checked to ensure the extension matches the actual file type, and are processed only inside the server, never sent out.
Accuracy — minimal hallucination
- ✓Rather than generating sentences, it answers only from verified Q&As and institutional material, minimizing hallucination
- ✓Every answer comes with an official source and a verification date
- ✓If no prepared answer exists, it says it doesn't know and suggests similar questions
- ✓If a question is ambiguous it asks back, and if there's a typo it corrects it with "Did you mean…?"
- ✓It understands follow-ups like "What about the fee, then?" together with the previous question
Security — data stays inside the institution
- ✓Installed on the institution's own server (on-premises)
- ✓The license key is verified by signature only, inside the server, without internet (offline verification · Ed25519 digital signature)
- ✓Never calls external AI or agent APIs
- ✓Works without an internet connection, so it can run on closed networks (offline installation bundle provided)
- ✓Questions and documents are not sent to us either
- ✓Question logs store no IP addresses or account information
Cost — no fees that grow with usage
- ✓Zero per-question token cost — usage fees don't rise as questions increase
- ✓No GPU required
- ✓License + annual maintenance, not a monthly subscription
- ✓Budgets can be planned accurately in advance
Customization — exactly how the institution wants it
- ✓Formal or polite speech style, emojis, answer length and forms of address are all configurable
- ✓Greetings, closings, "don't know" messages and contact information can be tailored too
- ✓Start from tone presets (courteous administrative · friendly counseling · concise business · plain Korean)
- ✓Preview a different tone instantly in the chatbot screen
- ✓White label — replace the product name, logo, colors, favicon, footer and answer-card name with the institution's own, and preview them
Operations — leave it to us and it keeps improving
- ✓Just give us your material — we handle the tuning
- ✓Evaluation questions check that different phrasings get the same answer
- ✓Unanswered questions are collected and filled in through regular tuning
- ✓Each question set is versioned, recording what changed and when
- ✓Setup wizard — after installation, set the license, institution details, operator account, chatbots, tone, languages and branding step by step in the web interface
- ✓The chatbot doesn't stop when the maintenance period ends — only new content deployment and upgrades wait until renewal
- ✓Institution dashboard — usage, page views (human · AI), FAQ additions, edit requests, document submission, monthly reports
- ✓Operator admin screen — FAQ/Q&A registration · tone · question logs · validation · deployment · rollback
- ✓Automated validation before deployment; deployments take effect immediately with no service interruption
Domain templates — start right away
- ✓Demos for 7 domains: local government civil services · university academic affairs · internal regulations · welfare · education office · library · foreign residents
- ✓Each domain has its own question set, so you can adopt only the domains you need
- ✓Connect to your website or internal systems via the web interface and API (JSON)
- ✓English by default — Korean, Chinese, Vietnamese, Japanese and Spanish as options
- ✓English, Chinese (Simplified), Vietnamese, Japanese and Spanish can be added as options
At a glance
| narubot™ | Typical generative AI chatbot (subscription) | |
|---|---|---|
| Basis of answers | Only verified Q&As and institutional material · sources shown | Model generates sentences · answers may be factually wrong |
| When it doesn't know | Says it doesn't know and suggests similar questions | May produce a plausible-sounding answer |
| Hallucination | Structurally minimized — uses verified answers only | Risk of plausible but factually wrong answers |
| External AI calls | None | An external model call per question is typical |
| Where data lives | Inside the institution's server (on-premises · closed networks possible) | Often sent to the vendor's cloud |
| Cost structure | License + annual maintenance · zero per question | Monthly subscription + usage-based fees |
| Hardware | Standard server · no GPU needed | Vendor servers or GPU servers |
| Tone | Tailored to the institution via settings · preview | Adjusted through prompts |
| Applying content changes | Validated and deployed from the admin screen · live immediately, no restart | Varies by product |
| License verification | Offline key verification — no internet connection needed | Connects to vendor servers to verify the account |
| Institution branding | White label — the institution's own name, logo and colors | Vendor branding is usually shown |
This comparison is based on the common setup of typical subscription-based generative AI chatbots and may vary by product.
How your documents become answers
We follow the same steps as RAG (retrieval-augmented generation) — ingest → chunk → index → retrieve → generate — with two differences: a human validation step before anything goes live, and replies are composed from verified answers with tone rules instead of being written by an LLM — so there are no unsourced answers or hallucinations, and it runs on closed networks without internet.
-
1UploadIngest
Manuals, regulations, notices, FAQ spreadsheets — stored on this server only; text extracted automatically
-
2Split into piecesChunking
Split by article, paragraph and table so each piece covers one topic
-
3Write Q&AKnowledge
For each piece: several ways people ask, a confirmed answer, official source and check date
-
4Index by wordsIndexing
Pre-built questions are split into words and linked to the concept dictionary — words with the same meaning (synonyms, abbreviations, other languages), e.g. passport = 护照 = パスポート
-
5Validate → publishNot in plain RAG
Human review → automatic validation against a question set → publish; roll back any time
-
6Retrieve → composeRetrieve · Generate
Match the question by words and concepts → confidence check → compose the reply from the verified answer with tone rules (greeting, register, closing, length) plus the source; unanswered questions feed the next tuning
VGA — Verified Grounded Answering
This is how narubot™ answers. After understanding the question, it looks for the answer in evidence the institution has verified, and only when it is confident does it show that answer together with the official source (the verification date is on the full answer page). It never makes up new sentences.
- 1Question200-character input box · previous question is used only for follow-ups
- 2Language detectionSelected language · requests like "answer in English" · the question's script (한국어 · English · 中文 · Tiếng Việt · 日本語 · Español) — answers only in enabled languages
- 3UnderstandingConcept matching (different words, same meaning) · follow-ups ("What about the cost?" → merged with the previous question) · typos · Korean/English keyboard-layout slips
- 4Finding evidenceOnly from verified material 🏢 Institution info
Hours · contacts · location📖 Glossary✅ Verified Q&A
Official source · verification date - 5Confidence checkMatch score between question and evidence · whether key concepts match · whether the meaning splits into more than one reading
- 1Question
- 2Document retrievalFinds document chunks similar to the question (usually by vector similarity)
- 3Generative model writes the textA language model writes a new answer using the retrieved chunks (external AI API or GPU server)
- 4AnswerIts strength is answering flexibly across a broad range of documents
Hallucination risk — If the retrieved chunks don't match the question or the model adds content, it can produce plausible but factually wrong sentences. That is why public-sector guidance needs additional review safeguards.
| VGA | RAG | |
|---|---|---|
| Answer text | Exactly as verified by the institution | Newly written by the model |
| When unsure | Says it doesn't know · suggests similar questions | Depends on model · settings |
| External AI calls | None (zero per question) | Usually used |
| How to fix | Edit the Q&A and deploy → live immediately | Adjust documents · prompts |
Legend
- Verified evidence — material confirmed by the institution (source · verification date)
- Judgment — confidence score · concept match
- Asking back — when the meaning splits or a typo is suspected
- Don't know — never makes things up when there's no evidence
The follow-up handling and typo correction in step 3 are a second search that runs when the first search result is weak.
VGA principles
- ① Answer text comes only from material verified by the institution — no new sentences are made up.
- ② A Q&A without a verification date cannot be deployed (automated validation). Answers show an official source link, and the full answer page shows the verification date.
- ③ If the meaning is ambiguous it asks back; if there's no evidence it says it doesn't know.
- ④ Unanswered questions are logged and filled in through regular tuning — question logs store no IP or account information.
Frequently asked questions about narubot™
What kind of chatbot is narubot™?
narubot™ is a chatbot for institutions that answers only from verified material. It is installed on the institution's server, and instead of calling third-party generative AI (ChatGPT, Claude, Gemini, etc.), its in-house engine produces answers inside the institution's server.
Without external AI APIs, how does it produce answers?
It uses VGA (Verified Grounded Answering). It understands the question, finds evidence in pre-verified Q&As and institutional material, and answers with the source and verification date attached. If there is no evidence, it doesn't make anything up — it says it doesn't know.
Doesn't it hallucinate?
Because it uses only verified answers rather than generating new sentences, hallucination is structurally minimized. If no prepared answer exists, it says it doesn't know and suggests similar questions.
Do questions or internal documents ever leave the institution?
No. All processing is completed inside the institution's server, and nothing is sent to external AI companies or to us. Question logs store no IP addresses or account information.
Can it be used on a closed network without internet access?
Yes. It works without an internet connection and can be installed with the closed-network installation bundle.
Does it need a GPU?
No. It runs on a standard server; the minimum specification for the Lite tier is 2 vCPU · 4GB memory · 50GB SSD.
How much does it cost?
It is licensed as one license per institution. Lite is $8,250, Standard $13,500, and Advanced from $22,500 (suggested retail); initial tuning and annual maintenance (15% of the license price per year, starting after acceptance) are charged separately. There is no per-question token cost. All prices exclude VAT, and contracts are made through partner companies.
Can it be procured through a direct (private) contract in Korea?
Yes. In Korea, all standard packages fall within the range where public institutions can sign a negotiated (direct) service contract — an estimated price of KRW 50 million or less, excluding VAT. The actual contract method follows each institution's procurement rules, and contracts are made through our partner companies.
Does it support other languages?
English is the default, and Korean, Chinese (Simplified), Vietnamese, Japanese and Spanish can be added as options, from $3,375 per language.
What does the institution need to do itself?
Simply upload or send the questions, FAQs, civil-service handbooks, regulation documents and other material you have through the institution dashboard. FAQ additions and edits can also be requested from the dashboard — we verify and apply them — and you can see progress, usage and monthly reports there too. We take care of answer tuning, tone customization, regular tuning and upgrades.
Can the tone be adapted to our institution?
Yes. Formal or polite speech style, emojis, answer length, forms of address, greetings and the "don't know" message are all configurable, and you can preview them in the chatbot screen.
Does the service stop when content is changed?
Changes to answers, FAQs and tone go live immediately without a restart once they pass automated checks, and can be rolled back in one step if there is a problem. We also make every effort to carry out engine upgrades and security updates without service interruption. However, a short maintenance window may be needed in some cases, such as database schema changes or server OS patches (reboots); in that case we notify you in advance and do it during low-usage hours.
What domains can it be used for?
We provide demos for 7 domains: local government civil-service guidance, university academic guidance, internal regulation search, welfare counseling, education office and school civil services, library services, and daily-life guidance for foreign residents. Other domains can also be set up based on your material.
How do installation and licensing work?
After installing on the institution's server with a single installation script, you set the license key, institution details, operator account, chatbots to use, tone, languages and branding step by step in the web-based setup wizard. The license key is verified by digital signature only, inside the server, without internet — so it works on closed networks too. Without a key, it runs as a trial with one chatbot and Korean only.
Does the chatbot stop when the maintenance period ends?
No. To keep public services uninterrupted, the chatbot keeps answering; only new content deployment and upgrades require entering a renewal key. A renewal notice is shown in the admin screen.
Can it use our institution's name and logo?
Yes (white label). In the admin screen you can change the product name, logo, colors, favicon, footer text and the name shown on answer cards, and preview them immediately. Hiding the 'Powered by narubot™' mark is available with a white-label license.
How do I request adoption or a demo?
Leave your institution name and area of interest on this site's Contact page (/contact), and we will get in touch.