Cuaba

Each question is classified — canonical policy questions route to the approved-document spine, ordinary questions to the book corpus — and the reply language is detected from the question itself.

The model must return discrete answer segments, each carrying explicit source IDs, because free-form prose cannot be checked and is not accepted.

A pure, unit-tested validator checks every citation against the sources actually retrieved, and a single unknown or missing citation rejects the entire draft rather than being patched locally.

A second model call re-checks grounding only where it earns its cost, on high-stakes questions and on retrieved book evidence, while answers resting purely on the approved spine skip it.

Vector search retrieves passages per book rather than globally, so each book answers from its own pages and the answer stays attributable to its source.

With several books answering independently, the system synthesises one recommendation across their competing perspectives instead of averaging those perspectives away beforehand.

A governed set of Markdown documents carries structured metadata — authority level, review date, version, language, aliases, keywords and a high-stakes flag — giving the system a tier it can treat as approved truth rather than retrieved text.

The governed documents are searched lexically before anything else, and a strong match answers immediately without spending an expensive retrieval pass.

A language edition is generated once, cached and resumable, so a language tab that has been opened before opens instantly.

Translation chunks pack whole paragraphs up to a size budget and fall back to sentence splitting only for a single oversized paragraph, which keeps blank lines intact all the way to the model.

Longer books are split into navigable chapters using real chapter titles detected from the text in all three languages rather than arbitrary page breaks.

A slug scheme of the form voice-male__booklang-es__title lets one pipeline serve every combination of book, chapter, passage, language and voice, each cached, shareable and independently addressable.

Narration runs through Microsoft's neural voices over the msedge-tts protocol, the same Azure Neural voice family commercial APIs bill for, so the 68 audiobooks generated to date cost nothing.

Because the free endpoint throttles bursts and drops sockets, synthesis is serialised through a queue, retried up to six times with exponential backoff and validated as non-empty, since a dropped socket produces truncated audio that would otherwise be cached and shared as complete.

Narration sits behind a provider-agnostic interface with an Edge, Gemini, OpenAI and ElevenLabs chain, so a long generation job survives the primary engine failing mid-run and the platform degrades rather than breaks.

Vector RPC calls carry an explicit 15-second timeout and return fewer or zero passages, a case every caller already handles, rather than hanging a serverless request until platform timeout.