Retrieval-augmented generation

Ask your documents.
Verify every answer.

Upload what you know. Ask in plain language, typed or out loud. Every sentence comes back with a citation you can open and read — grounded in your documents, or not answered at all.

An example answer and the passages it was built from.
01

Ingest

Paste text or upload a PDF. It is split into overlapping chunks and embedded with text-embedding-3-small.

02

Retrieve

Your question is embedded too, then matched against your chunks by cosine similarity over an HNSW index in pgvector.

03

Ground

Only the top-k passages reach the model, and it must cite them. No supporting passage, no answer.

Voice sits either side of those three steps rather than inside them. A spoken question becomes text before anything is searched, and a finished answer is read back sentence by sentence as it streams — so the pipeline never learns which one you used.