SIRANA
Emergency regulation is exactly the kind of material where a confidently wrong answer causes harm. SIRANA lets people ask about emergency regulations in plain language, and answers from the real regulation and education documents rather than from the model's memory.
- Role
- AI Integration Engineer, Frontend Engineer
- Discipline
- NLP, Public information
Built with
- Next.js
- NLP
- LLM
- RAG
- Vector store
A general-purpose model asked about emergency regulation will produce something plausible whether or not it knows the answer. For this subject that is a failure mode, not a quirk.
So answers are retrieved (RAG) from the regulation and education documents before the model writes anything. The response is anchored to source material that exists, which also means a reader can go and check it.
The web interface
Built the frontend in Next.js, with the chat experience at the centre of it.
The chat flow and prompt design
How a question is framed, what context is retrieved, and how the answer comes back.
Interface ⇄ AI service
The connection between the frontend and the AI service behind it, including the retrieval step over the vector store.
Next project