An AI guide grounded in

An AI guide grounded in 56 texts of conscience. It quotes only what it finds, shows only the sources it used, and finds the right one 94% of the time. The brief

VoxAnima is a conversational guide for people facing a moral choice at work: the moment you see something wrong and have to decide whether to keep going. It speaks in the voice of its founder, an activist and author with 35 years in the work. It draws on a fixed library of nonviolence and wisdom literature: Gandhi, the Gospels, the Hebrew Bible, the Qur'an, the Dhammapada, Tolstoy, the Quakers, the Stoics and more.

One rule mattered more than anything else. Every claim had to come from a passage the system actually found, and the source line under each answer had to be true.

What I built

A searchable library of 56 texts, split into 1,600+ passages, stored in Postgres with pgvector and searched by meaning on every question.

Chunking that respects the texts. Scripture is split by whole verses, so no passage cuts a verse in half or crosses a chapter. Each passage carries its own reference, like Isaiah 58:1-14, so the answer can say exactly where a line came from.

A search cutoff measured, not guessed. The original spec's threshold was tuned for a different embedding model and returned nothing. I measured the real score range and set the cutoff from the data.

A balance rule. Scripture is about 60% of the library, so it used to crowd out every other voice. Each answer now takes at most 2 passages from any one source, which leaves room for Woolman, Tolstoy or the Buddha to speak.

Sources you can trust. The model numbers each passage it uses. The app shows a source line only for passages the answer actually drew on. A passage that was found but not used gets no byline.

A test set for search. 34 real questions with known answers, run before and after every change. The right source lands in the top 5 for 94% of them.

Fail loudly, never silently. When the database went down in production, the app had kept answering with no sources and nobody could tell. I made failures visible, added a daily health check, and the test suite now refuses to pass against a dead backend.

Results

94% of test questions find the right source in the top 5 (MRR 0.78)

Only cited passages get a byline: in the cover's live example, 5 passages were found, 3 were used, and 2 got no byline

The founder's voice comes from his own recorded answers, held as a reference set the model writes against

Stack

Next.js · Supabase (Postgres + pgvector) · Voyage embeddings · Claude via the Vercel AI SDK · Vercel