marola.knowledge

The corpus is a directory of Markdown files (knowledge/ at the repo root by default), plus a safety/ subdirectory (MIP-0022) for documents whose answers must always carry the emergency footer.

One indexable unit: a paragraph-ish slice of a corpus document, with the document's provenance.

Turns text into vectors.

Local-first RAG index: embeds every corpus chunk once, stores the vectors as JSON under ./data/, and re-embeds only when the corpus or the embedding model changes (a fingerprint of file names, sizes, mtimes and the model name).

Retrieval over marola's curated marine-knowledge corpus (the Markdown files under knowledge/) — the RAG half of FUTURE-WORK.md §9.1, local-first.

"Ask the ocean": retrieve the top passages for a question, then have the local model answer only from them, citing [n].

A retrieved chunk with its provenance — source is the URL the corpus document cites.

The emergency footer marola appends to any answer grounded in a knowledge/safety/ document (MIP-0022).