---
title: "The AI Behind This Site, Explained"
description: "Plain-English explanations of the AI behind Truth Promoters — speech to text, translation, synthetic voice, phonetic transliteration, how passages are found, image generation and drafting."
canonical_url: "https://www.truthpromoters.com/help/ai-explained"
type: "help"
section: "Help Center"
keywords: "what is RAG, what is an embedding, what is a reranker, speech to text, text to speech, transliteration"
---

# The AI Behind This Site, Explained

This page explains, in plain English, what each of the AI systems behind this site actually does.
It is a companion to [Working With AI](/help/working-with-ai), which explains *why* we do this at
all and what a person still decides.

> [!NOTE]
> None of this requires you to know anything technical. If a term below is unfamiliar, that is
> exactly why the page exists — the jargon is other people's, not yours.

## Speech to text

**What it is:** software that listens to a recording and writes down what was said.

Sometimes called *transcription* or *ASR*. Give it an hour of a lecture and it returns an hour of
text, with timings, so a particular sentence can be found and played.

**What we do with it:** lectures, sessions and talks become searchable text. Once something is
text, it can be quoted, translated, indexed and read — and a two-hour recording stops being an
opaque blob you have to sit through to find one paragraph.

**What it gets wrong:** names, unusual terms, and anything said over background noise. A person
checks it against the audio.

## Translation

**What it is:** the obvious thing — moving text from one language to another.

**What we do with it:** a first pass on Persian and Arabic material, which a person then works
over. It is genuinely good at ordinary prose and genuinely unreliable at exactly the material we
care about most: technical terms, honorifics, and anything where a word carries a specific
scholarly meaning.

**The harder problem, which is not really translation at all:** an idiom that is transparent to
an Iranian audience can be opaque or misleading to a British one. Nothing is mistranslated; the
weight simply does not survive the trip. This is an open problem for us and we say more about it
on the other page.

## Text to speech

**What it is:** software that reads text aloud in a synthetic voice.

**What we do with it:** every substantial page on this site — including this one — can be
listened to rather than read. That matters for people who commute, who have poor eyesight, or who
simply take things in better by ear.

## Phonetic transliteration, and why it deserves its own section

**What it is:** rewriting Arabic, Farsi or Hebrew into letters that describe how the words
*sound*, so a synthetic voice can pronounce them.

**Why it is not a small detail:** a voice reading unconverted Arabic does not produce accented
Arabic. It produces noise, or worse, it produces something that sounds like words and is not. For
a site that quotes scripture and narrations, that is not a rough edge — it is the thing itself
being damaged.

So the conversions are **stored, reviewable and hand-correctable**, rather than being redone
invisibly on every playback. A conversion nobody has reviewed yet says so out loud before the
audio starts. And the model that performs the conversion cannot be swapped without first passing
a set of deliberately difficult passages.

## Retrieval — how we find the right passage

This is the part most worth understanding, because it is where a site like ours either earns
trust or loses it. Broken into its actual steps:

### Ingest

**What it is:** taking a book — often a very large one — and loading it into a form a computer can
search meaningfully.

That is not just copying the text in. Each work has to be divided into passages of a sensible
size, tagged with what kind of text it is and which tradition it belongs to, and checked for the
debris that real-world source files are full of: fragments, duplicated headers, broken encodings,
entries that are a chapter heading and nothing else.

**Why the size matters:** too long and only the opening ever gets read; too short and the passage
is so generic it weakly matches everything and grounds nothing. Where a text can be divided
cleanly we do; where it cannot — some vocalised Arabic carries no sentence marks at all, so
cutting by counting characters would sever words — **we leave it whole and record why**.

### Embedding

**What it is:** turning a passage of text into a long list of numbers that represents its
*meaning*, so that two passages about patience end up near each other even when they share no
words at all.

This is the trick that makes searching by idea rather than by keyword possible. A keyword search
for "patience" misses a narration about enduring hardship that never uses the word.

**The catch:** it is approximate. Two passages can land close together for reasons that are real
but useless — sharing a register, or a formula, rather than a subject.

### Search and reranking

**What it is:** two stages, and the second one exists because the first is fast rather than
careful.

The first stage sweeps the whole library and returns a shortlist of plausible candidates,
quickly. The second — the **reranker** — reads each candidate *against the actual question* and
scores how well it genuinely answers it. It is much slower per passage, which is why it only ever
sees the shortlist, and much better at telling apart "shares vocabulary with the question" from
"is actually relevant to it".

That distinction is not academic. A passage can score well on the first stage purely by sharing
subject matter and say nothing whatsoever to the point. Removing the second stage roughly halves
how often the most relevant passage comes out on top.

### Why this whole apparatus exists

So that when something is quoted here, **it was found in a real book rather than composed to
sound right.**

A language model asked to supply a supporting narration will supply one. It will be fluent, well
formed, correctly styled and sometimes wholly invented. The defence against that is not a better
prompt. It is giving the model an actual library to draw from, requiring it to say when it has
found nothing, and having a person check that the references resolve.

## Image generation

**What it is:** producing an illustration from a written description.

**What we do with it:** artwork for Reflections and the panels in Vignettes. **We do not generate
depictions of the Prophet, the Imams, or the Ahl al-Bayt (AS)**, and we do not use it to
manufacture photographs of real events.

## Drafting

**What it is:** a model producing a first version of a piece of writing from a brief.

**What we do with it:** first drafts of Reflections, Vignette stories and event write-ups, always
worked from material retrieved from our own library rather than from whatever the model
remembers.

A first draft is where these systems are most useful and least trustworthy at the same time. They
are fast, fluent, and perfectly willing to be confidently wrong, so what they produce is treated
as raw material for a person rather than as output.

## What is not automated

Worth stating plainly, because a page listing capabilities reads as though a machine writes the
site:

- **Nothing publishes itself.** A person schedules it.
- **Overnight generation is off by default.** A day nobody writes for is a blank day.
- **What a session is about, and what it argues**, is decided by people who have studied.
- **Whether a source may be cited at all** is a decision made in advance by people, not per
  request by a model.
- **Corrections** — to a transliteration, a translation, a reference — are made by a person and
  stored, so the same mistake is not made again tomorrow.

---

If you want the reasoning behind all of this rather than the mechanics, that is on
[Working With AI](/help/working-with-ai).

## Related topics

- [Working With AI](https://www.truthpromoters.com/help/working-with-ai) — Why we work with AI rather than use it, what a person still decides, and where we refuse
- [Listening to the Site](https://www.truthpromoters.com/help/listening) — Play articles and recordings, build a queue, and pick up where you left off