Case Study · Cultural Heritage

Building an AI Oral History Archive: RAG for 60 Hours of Testimony

An African cultural heritage institution asked for a way to make more than 60 hours of oral testimony searchable and citable, without flattening the material into a generic chatbot. This is how I built it.

By Brandon Da Costa, Founder, Digiton Dynamics

The material

The institution held more than 60 hours of recorded oral testimony and over 1,000 pages of transcripts documenting cultural heritage: memory, history and testimony that exists nowhere else in that form. The brief was not 'build a chatbot.' It was: make this material findable and usable by researchers and the public, without losing the nuance, the disagreement between accounts, or the fact that oral testimony is not a single clean source of truth the way a technical manual is.

Why a general-purpose model was never going to be enough

Ask a raw language model a question about this kind of archive and it will answer from whatever it happened to absorb during training, which is either nothing about this specific material, or worse, a plausible-sounding invention. Heritage and testimony work cannot tolerate that. Every answer has to trace back to an actual transcript, an actual timestamp, an actual speaker. That requirement shaped the entire architecture.

The transcription and indexing pipeline

Before any retrieval could happen, the raw testimony had to become structured, searchable text: transcribing over 60 hours of recorded material, then breaking those transcripts into passages small enough to retrieve precisely but large enough to preserve context, since a fragment stripped of its surrounding conversation can misrepresent what a speaker actually meant. Each passage keeps metadata about who is speaking, when, and in what part of the recording, so a researcher can go from a retrieved answer straight back to the original audio if they need to verify it themselves. That last step, the ability to go back to the source recording and not just a transcript, was something the institution specifically asked for, and it shaped how I structured the indexing from the start.

The RAG pipeline

The transcripts were processed and indexed so they could be retrieved by meaning, not just by keyword, since testimony rarely uses the exact phrasing a researcher searches for. Each retrieved passage carries its provenance: which recording, which section, which speaker, so an answer is never presented as a free-floating fact. It is presented with its source attached, the way a historian would expect to see a citation. That traceability was the actual engineering goal, not a nice-to-have layered on afterward.

Handling disagreement, not smoothing it over

Oral history is full of accounts that do not agree with each other, and that disagreement is often the most historically important part of the material, not a bug to be resolved. A naive RAG system tends to synthesize a single confident answer and quietly discard the tension between sources. I built the retrieval and presentation layer to surface multiple relevant passages when accounts diverge, rather than collapsing them into one voice. The system's job is to make the material navigable, not to decide which memory is correct.

Presenting it publicly

I presented this work publicly at Universidade Aberta in Lisbon in April 2026, walking through the RAG pipeline design and the specific challenge of grounding AI in historical testimony rather than technical documentation. That distinction matters more than it sounds: most RAG writing assumes source material that is internally consistent. Heritage archives are the opposite, and the architecture has to be honest about that instead of pretending otherwise.

What I deliberately did not build

I want to be precise about what this project is, because it would be easy to overstate. This was work for a cultural heritage institution, not a government agency, and it was a defined engagement, not an ongoing operational relationship. I mention that because heritage and historical work attracts a certain kind of inflated claim, and I would rather be exact about the scope than borrow credibility the project did not actually establish. What I can say plainly is that the platform works, the transcripts are searchable and citable, and the architecture holds up under the specific demands of testimony-based material.

What carries over to other work

The core lesson generalizes well beyond heritage: whenever the source material itself is contested, incomplete or multi-voiced, a RAG system that forces a single confident answer is doing something dishonest, even if it reads well. Good architecture surfaces the sources and lets the human weigh them. That principle now shapes every knowledge system I design, not just this one. If your organization is sitting on an archive, a body of documentation or institutional knowledge that deserves better than a filing cabinet, that is exactly the kind of AI system I like building. Get in touch if that sounds like your problem.

Frequently asked questions

Related