Skip to main content
FirstCoast.ai
All case studies
Case study: Cultural heritage archive

AI Chat Agent for a Historic Art Archive

The complete catalogue of a nineteenth-century portrait painter, spanning 238 works, 70 archival photographs, and a translated diary, rebuilt as a modern archive with a grounded AI chat agent as its front door and a standing invitation to help recover lost works.

How the grounded archive chat agent worksA nineteenth-century portrait painter's complete archive, rebuilt as a structured content system: the biography, the catalogue of all 238 works, 217 dated diary entries, and 70 archival photographs, with twenty years of legacy URLs preserved through the rebuild and structured metadata on every work so the collection stays legible to search and answer engines. Two processes touch that archive, and they are deliberately different. The chat agent only reads. A visitor asks a question in the site chat, and the agent answers through a two-layer prompt: the persona and the guardrails sit on top, where valuations and authentication opinions are refused outright, and the knowledge base assembled from the archive sits underneath them, treated as data the model reads and never as instructions it follows, which is what stops a prompt injection hidden in page content. Conversations cost about four cents each, putting the whole public-facing agent at ten to twenty-five dollars a month. Three kinds of answer come out. It cites the specific work it references, so every answer is traceable back to the archive. Where the archive does not record something it says so, rather than inventing an answer. And because many works are lost and known only from photographs, a visitor who knows where one went is asked to contact the family, which turns every conversation into a possible recovery. The second process is the image pipeline, and it is the only thing that writes. Scans of original artwork have the frame baked in, so AI proposes a crop but never applies one: every crop waits on a recorded human decision, and only an approved crop goes back into the archive. The archival photographs of the lost works are never cropped at all.THE ARCHIVE OF RECORDSTRUCTURED CONTENTBiography238 works217 diary entries70 archival photographsStructured metadata on every worklegible to search and answer enginestwenty years of legacy URLs preservedA VISITOR ASKSSITE CHAT“Who is the woman in this portrait, and where is it now?”The chat agenta two-layer prompt, guardrails on topFour cents a conversation$10 to $25 a monthPERSONA AND GUARDRAILSNo valuationsNo authentication opinionsKNOWLEDGE BASE, FROM THE ARCHIVESITE CONTENT IS DATA, NEVER INSTRUCTIONSCites the workthe specific one itreferencesTRACEABLE TO THE ARCHIVEOr admits a gapwhen the archive does notrecord something, neveran invented answerLOST WORKSContact the familyif a visitor knowswhere a lost work wentTHE IRREPLACEABLE SCANSScanned original artworkwith the frame baked inAI proposes a cropit never applies oneHUMAN GATEA recorded human decisionbefore any crop is appliedNO CROP WITHOUT ONENEVER CROPPED AT ALLPhotographs of lost works

The challenge

  • A decade-old website held the archive but could not answer a visitor's questions about the works, the artist, or the history
  • Many works are lost, known only from archival photographs, and every visitor is a potential lead in recovering one
  • Scans of original artwork are irreplaceable; nothing automated could be allowed to alter them unsupervised

What we built

A rebuilt archive site with an AI chat agent grounded strictly in the archive itself: the biography, the catalogue of all 238 works, the diary, and the photographs.

  • Answers only from the archive and cites the specific work it references; when the archive does not record something, it says so instead of inventing
  • Never offers valuations or authentication opinions, and asks anyone with knowledge of a lost painting to contact the family
  • An image pipeline proposes crops of baked-in frames on scanned artwork, but every crop requires a recorded human decision, and archival photographs of lost works are never cropped at all

How it's built

A structured content system holding the full catalogue, and a chat service with a two-layer prompt: persona and guardrails above a knowledge base assembled from the archive, with prompt-injection defenses that treat site content as data, never as instructions. Structured metadata on every work makes the archive legible to search and answer engines.

Results

  • 238 works, 217 dated diary entries, and twenty years of legacy URLs preserved through the rebuild
  • Conversations cost about four cents each, putting a public-facing agent at roughly $10 to $25 a month
  • Every answer is traceable to the archive, so the family's scholarship stays the single source of truth

What knowledge is locked in your organization's archive?

A grounded chat agent turns a static collection into a conversation, without ever speaking beyond the record.