How sourced answers work
The model only works inside the current collection. It searches the documents in that collection, writes an answer, and points back to the source it used. If the answer looks wrong, incomplete, or too broad, you can inspect the original reference instead of trusting the model blindly.
That is the difference between grounded answers and open-ended AI chat. A search engine returns whatever ranks. A general chatbot gives you a polished answer with little visibility into where it came from. Disaster Clippy keeps the document in view so the answer remains tied to something you can verify.
What the collection is
Narrow and legible beats wide and opaque. Disaster Clippy searches a fixed collection you can inspect. You can see what is being searched, choose which sources to include, and know that nothing outside those boundaries is influencing the answer.
Why hosted and local both matter
The hosted app is the easiest way to evaluate the model. The local runtime is what makes it useful on your own terms: on your hardware, with your data, and with the option to run fully offline with no external APIs and no data leaving your machine.
How the builder path works
Full control belongs with the people who understand their sources. Scraping, transcript generation, validation, and pack building stay in the local admin path. That keeps the public demo simple while giving builders and maintainers the full system for creating their own collection.
Who built this
Disaster Clippy is built by Bryan Hellard. He has a background in Robotics and Systems Engineering and has spent the past few years working on disaster impacts and response. His goal is to make more data accessible so people can make informed decisions for themselves.