How a search works
When you ask a question, the system embeds it as a vector and searches the collection for semantically relevant documents. It then uses a language model to synthesize an answer from those documents and returns the answer alongside citations pointing back to the original sources. It does not search the open web and does not generate information outside what is in the collection.
What the collection is
The collection is made up of packs - discrete bundles of knowledge, each covering one body of content. Each pack includes source text, vector embeddings, metadata, and optionally offline backups of the original pages. Publicly, you experience that as a collection you can inspect and filter. You can see which sources are loaded, choose which ones to include in a search, and trust that nothing outside your selection influences the answer.
The three layers
The hosted app is the public-facing surface: search and chat without any setup. The local runtime is for self-hosted deployments, offline use, and people who want control over what goes in the collection. The advanced local admin toolkit is for building and curating sources - scraping, indexing, translation, and packaging.
These are not three different products. The hosted app and a local deployment run the same code pointed at different collections. Packs are what move between them.
How the .com site fits in
Everything you need to run the system - the runtime, the ingestion tools, the admin panel - is in the public repository.
Public repository on GitHub