Case study · Documents & language
Korean insurance chatbot & voice assistant
A retrieval-augmented chatbot that answers customer questions from the company’s own insurance documentation in Korean (compared head-to-head with ChatGPT) and a speech pipeline that lets customers ask the same questions by phone.
Business result
Customers get answers specific to this insurer’s products and procedures, by chat or by phone, in Korean, instead of the generic answers a general chatbot gives.
Client type
Client project for a South Korean company
Where this applies for you
Any regulated business whose answers must come from its own documents: insurance, banking, healthcare, legal.
Technical result
- 한국어Korean end to endQuestions, source documents and answers all in Korean
- 2channels, one engineChat and phone share the same retrieval and generation
- 7step call flowGreeting, policy check, spoken question, spoken answer
The problem
General-purpose chatbots answer insurance questions fluently but generically: they describe what insurers usually ask for, not what this company actually requires. Customers needed answers specific to the company’s products and procedures, in Korean, and often over the phone rather than in a chat window.
What we built
We indexed the company’s insurance documentation and built a RAG chatbot that answers from retrieved passages, then evaluated it against ChatGPT on the same customer questions. For the phone channel, a speech-to-text and text-to-speech pipeline wraps the same engine: the system greets the caller, verifies the policy, takes the question by voice and reads the grounded answer back.
Workflow
How it works, step by step.
The pipeline in 5 stages, from index to respond. Each stage has a clear input, a clear output and a reason to exist.
- 01
Index
Insurance documentation (products, procedures, required documents) chunked and embedded for retrieval.
- 02
Ask
Customers type a question, or speak it on a call; speech-to-text converts it to text.
- 03
Retrieve
The most relevant passages from the company’s documents are pulled for each question.
- 04
Answer
The model writes a Korean answer from those passages: inquiries, policy details and claim-related questions.
- 05
Respond
The answer is shown in chat, or synthesized to speech in real time for the caller.
Results
What it produced.
Same question, two systems
Our RAG chatbot · grounded in company documents
- 신분증 (주민등록증, 운전면허증 등)
- 가족관계등록부 (가족관계증명서)
- 주민등록등본
- 기타 필요한 서류 (예: 사고증명서, 경찰신고서류 등)
The company’s own checklist: ID card, family-relation certificate, resident-registration abstract, and case-specific documents such as accident or police reports.
ChatGPT · general model
- 보험청약서 (회사양식)
- 신분증 (사진이 부착된 정부기관 발행 신분증)
- 기타 회사가 요구하는 증빙서류
A generic answer (an application form, ID, “other documents the company requires”) plus a caveat that requirements vary by insurer.
The RAG answer reflects what this company actually asks for; the general model can only describe what insurers usually ask for.
Voice channel · call flow
- 1System · TTS보험에 가입하셨나요? 네 또는 아니요로 답해주세요.Do you have a policy with us? Please answer yes or no.
- Caller · STT네.Yes.2
- 3System · TTS보험 아이디를 알려주세요.Please tell me your policy ID.
- Caller · STT[Policy ID, spoken]4
- 5System · TTS고객님 질문을 입력해 주세요.Please tell us your question.
- Caller · STT제 보험에 대해 알려주세요.Tell me about my insurance.6
- 7System · RAG → TTS저희는 주택화재보험을 제공하며, 이는 주택 화재로 인한 손해에 대해 보장합니다…Reads back a policy summary generated from the retrieved documents: coverage, how the insured amount is set, how claims are calculated.
Stack & techniques
- Python
- Retrieval-augmented generation
- Vector search
- Korean NLP
- Speech-to-text
- Text-to-speech
Delivered by members of the Saerosense founding team. Client names are withheld unless public. References are available on request.
Start a project
Have a similar problem?
Send a few lines about your data and the outcome you want. We reply within 1 business day with an honest read and a proposed scoping plan.