From Critical Edition to Cited Answer

Station 1 — The source itself, not a copy of it Books enter from the critically edited digital editions of the Ahlulbayt Library, with their original structural markup: page breaks, footnote blocks, tables of contents, and the complete book card (author and death year, editor, edition, publisher, volume). The card is parsed automatically from the file itself, so citation data is never typed in by hand.

Station 2 — Structural ingestion that preserves the book’s fabric Every numbered footnote is merged into the body text at its exact anchor, tagged; a narrator’s biography stays attached to its place in the chain of transmission wherever the text is later split. Editors’ unanchored remarks are tagged separately, commentary blocks likewise, and the full hierarchy — book, chapters, sections — is rebuilt, including headings that appear only in the table of contents.

Station 3 — Semantic chunking that carries its own citation Text is split at meaning boundaries, and every passage carries in its metadata: book, author, volume, page number, its heading path, and its nature (body text, footnoted, commentary, editor’s note). Citation is not a garnish added at display time — it is a structural property of every passage from birth.

Station 4 — Dual retrieval, then reranking Two engines together: a semantic one that understands meaning across phrasings, and a lexical Arabic one that pins down the exact scholastic term — fused by reciprocal rank, then refined by a state-of-the-art reranker. A text is lost neither to a colloquial question nor to a rare term.

Station 5 — When no text exists: honesty, not invention The corpus does not compose, infer or patch. When nothing relevant exists it says so plainly: “No relevant results in the Moheet corpus.” A corpus that knows its limits is more trustworthy than one that answers everything.