Let a reader ask forwhat they cannot name.

It gives a patron a way into the material by conversation, so keyword search stops being the only door. You curate a set of records from the collection, each carrying its own metadata — creator, date, document type, holding repository, accession number, and whether it is first-hand or a later account. A reader asks a question in their own words, and the answer comes back with the passage and a route to the scan it came from. Material that would otherwise need a reader to already know what to search for becomes reachable.

No, and it is not designed to. Everything a patron can reach is a set of records your staff selected, described, and released. What it changes is the volume of first-contact questions that never reach a reference desk at all, either because the reading room is closed or because the reader did not know the collection existed. Staff time moves toward the questions that genuinely need an archivist.

By never letting the AI write a quote at all. It marks which part of which record it means, and our system copies the words straight out of the document, throwing away anything that does not match letter for letter. Across the 1,323 quotes in our published Humy 2.6 test, none had been reworded, none highlighted the wrong passage, and none came from anything but that figure's own records. One citation out of the 1,323 opened no source, and we report that rather than round it away.

Saying so is the behaviour we test hardest. One of the five reasoning categories in our 2.6 benchmark is exactly this: recognising an unrecorded fact, a false premise, or an apocryphal quotation and declining to fill the gap. Grounding helped most there, moving from 24.6% of checks passed without an archive to 59.3% with one. A confident answer to a question the records do not support is the failure mode this whole design exists to catch.

No, and the distinction matters for a repository. The quotation is compiled from your document and cannot be fabricated. The sentences around it are model-generated, and in the 2.6 run roughly one answer in eight still contained an invented date or document reference. We therefore put the record one click away from every claim, so a reader can check rather than trust.

Up to 50 source records per figure and up to 1,000 pages per record: scanned books and periodicals, correspondence, catalogue and archive records, photographs, paintings, audio recordings, and 3D-scanned objects. Pages with no text layer are transcribed one at a time by Gemini 3.1 Pro under an archival transcription prompt, which makes copperplate and secretary hands searchable. You approve an estimated cost before any page is processed.

You do, record by record. Each source can stay private to your institution, be shared across it, or be published openly. Restricted, embargoed, or culturally sensitive material simply stays out of the set you release, and because a figure can only answer from records you released, exclusion is enforced by what the archive contains. No model is asked to keep a secret.

No. Access can run through a plain link or a QR code with no account, email, or password, which suits a catalogue page, a reading-room sign, or an exhibition panel. Where you want conversations attributable, named entry can be required on a particular link instead.

Yes. At National Louis University in Chicago we built a figure of the university's founder, Elizabeth Harrison, from 25 of her writings, photographs, and documents supplied by the university library. Collection-based builds are also running with the University of Wisconsin–River Falls and the Private University of Education, Diocese of Linz in Austria. These are scoped with our team rather than self-serve.

With one figure and a small, well-described set of records. The whole repository can wait. Tell us the collection you would start with and three documents that sit at the heart of it, at hello@humy.ai, and we will scope a pilot against it.