La Diaria: The archive tool that failed, and what it teaches about buying
Six months, an outside consultant, and a RAG agent that lost to Google search — a governance failure more than a technical one.
Problem statement
La Diaria's first five years or so (roughly 2006–2011) were print-only and survive only as scanned PDFs, invisible to the paper's website and to search. With the paper's 20th anniversary approaching, the newsroom wanted to convert that archive into searchable data using RAG — retrieval-augmented generation, meaning an AI that can search and answer from your own documents — so that editors could plan coverage against two decades of prior reporting (for instance, drawing on years of euthanasia-law coverage when the law finally passed), and so that subscribers could eventually explore the paper's history.
The solution, as attempted
This was a buy, not a build. An external IT consultant was hired on the recommendation of senior newsroom leadership — a top-down hire — and funded for six months. The internal IT team had been asked first and candidly declined: they didn't know RAG, would have to study it, and lacked the time. The two staff members running the project set clear expectations: defined steps, a six-month timeline, and an offer of weekly or even twice-weekly meetings. The consultant called it an "easy-easy project" he would finish in the first month. A programme mentor provided close oversight, tracking milestones week to week. What was eventually delivered was a ChatGPT-based conversational agent over the archive.
Results
The main objective was not achieved. The delivered product contained many mistakes, and the final agent was less effective than a Google search — typing "La Diaria euthanasia" into Google gave better results than the tool built for the purpose. Along the way, the consultant repeatedly reported that everything was working and asked for more money and tools; the newsroom applied for and contributed additional internal funds, which were granted faster than the team's escalating complaints were addressed. By the newsroom's account, the mentor was the first to spot the problem, warning after three meetings in one week with no progress that the consultant might be playing them. Escalation then failed for months: the mid-level staff running the project called, emailed and visited leadership, who eventually acknowledged "you were right all along" — too late. The interviewee attributes the failure primarily to a communication and governance problem — no clear channel for raising the alarm — compounded by the project not being a leadership priority in a busy anniversary period. The morale cost was real: he describes feeling unbacked by the institution, doubting his own feedback, and being ashamed of having requested internal funds that were wasted.
Insights
The tool was never adopted into practice. The newsroom is preparing a retrospective presentation, built with its IT team, explicitly marking where the project failed.
Cost
Directly, six months of consultant funding from programme funds, plus additional money from the newsroom's own budget at the consultant's request. Indirectly, six months of sustained effort by two staff members on top of their day jobs, including weekly-or-more meetings, plus mentor time and the effort spent escalating.
Replicability potential
The transferable value is the failure lessons, which the newsroom articulates itself — and they are about buying, not about the technology. First, get a trusted intermediary to vet or recommend external developers rather than relying on internal seniority-based recommendations: newsrooms know how to judge journalists, not IT contractors. Second, tie payment to verified milestones — milestones existed here, but nothing happened when they were missed. Third, establish a clear internal escalation channel before the project starts, so mid-level staff raising alarms are heard. Fourth, beware the "easy project, done in a month" pitch on genuinely complex work.