OneSoul Wisdom

The problem

The public site did its job: teachers, guidance, contact. The harder need was private — a study library of classical Hebrew and Aramaic works that students could actually read, search and hear, in the language they think in.

Translating a sacred corpus is not a job you hand to a general translation engine. A term rendered loosely in one passage and differently in the next destroys the thing that makes the text usable for study. And a library nobody can search is a shelf, not a tool.

What we built

• A translation pipeline with a controlled glossary, so a term of art lands the same way on page 3 and page 300. English and Spanish are both carried from the Hebrew; Spanish additionally sees the aligned English as a terminology reference, never as its source.
• Paragraph-level full-text search on D1 using FTS5 — one indexed row per paragraph per language per book, so a reader finds the passage, not merely the chapter.
• Generated narration for every block, one small audio file at a time, served from R2. One recent addition alone produced 708 narration files per language.
• Session-gated access. Every reader endpoint returns 401 unauthenticated; the session is an HMAC-signed cookie, and the model API key never leaves the Worker — even the translation tooling calls through a gated proxy rather than holding a key locally.
• A resumable pipeline. Every stage — fetch, translate, upload, index, narrate — can stop and continue, because a corpus-scale job that cannot resume is a job that never finishes.

What it means

The most recent work added 88 sections across Hebrew, English and Spanish — 620 paragraphs and 708 narration files in each language — for roughly five dollars in translation cost, because the narration is generated locally and free. The reader is the newest thing running in our entire estate, and the only place we use D1 as a search engine rather than a record store.

Stack

Workers, D1 full-text search, KV, R2, AI translation, Text-to-Speech