System
Ask
Questions answered out of the project's own corpus, with the source attached. There is no generation step, so there is nothing that could invent an answer.
Why an answer here is a list of passages
A model that writes a paragraph over retrieved text can add a sentence the text does not support, and that sentence arrives wearing the same citation as the rest. Until Phase C adds a model that is forbidden from doing so, an answer is the passages themselves — less satisfying, and incapable of fabricating.
Retrieval is BM25 over the project’s own corpus. It is not a placeholder: it is the half of a hybrid retriever that needs no model, it runs on a free instance with no embedder, and it makes the citation contract testable before any inference is involved. The A9 spike chose bge-m3 for the vector half — see docs/models.md.