AskDiana Shows Its Working

There is a particular flavour of confidence that only a large language model can produce: the tone of a man who has never once in his life been wrong, delivered by a system that is, on a not-insignificant percentage of occasions, extremely wrong. Ask a frontier chatbot a question about your company's Q3 numbers and it will answer with the serene certainty of someone reading off a stone tablet, rather than someone guessing, plausibly, from a training cutoff that predates your Q3 by several years and your company entirely.

AskDiana was built on the mildly heretical idea that an AI system answering questions about your business ought to actually know something about your business, and ought to be honest about how sure it is when it doesn't. It sounds obvious. It is astonishing how few products manage it.

 

What It Actually Is

Strip away the marketing and AskDiana is a multi-tenant "Business IQ" platform. Organisations upload their documents, connect their ERP, and get a private subdomain (yourcompany.askdiana.ai) where staff can chat with the actual contents of the business: PDFs, spreadsheets, sales figures, inventory, finance data, the lot. It talks to Odoo and to Galaxy ERP, which is, and I say this with no particular enthusiasm for the phrase "number one," the leading ERP system across Greece and the Balkans, currently deployed via AskDiana for Sierra Studios. It has also been quietly running, in a rather more careful configuration, for the Refugee Council of Lithuania, where the requirements shift from "answer the question" to "answer the question without logging who asked it," and the platform pseudonymises IP addresses with HMAC-SHA256 and skips the login screen entirely, because for that particular user base, a login screen is not a convenience, it is a liability.

Under the bonnet, retrieval isn't a single vector search hoping for the best. It's hybrid: dense vector search, BM42 keyword matching, and a Neo4j knowledge graph, cross-checking each other before anything reaches the model. There's a semantic cache sitting in front of all of it, so the fortieth person to ask "what were our returns like in March" doesn't wait for a fresh round trip through the whole retrieval pipeline; if a near-identical question has already been answered above a 0.97 similarity threshold, the cached answer comes back instantly, and it's still correct, because the documents underneath it haven't changed.

 

The Bit Where It Checks Its Own Homework

Retrieval solves "does the model have the right information in front of it." It does not solve "will the model faithfully report that information rather than making something up that sounds equally plausible." For that, AskDiana can hand the final answer generation over to a sister system called Genius², which is best described as a small, disagreeable parliament of language models rather than a single overconfident narrator.

Genius² fires the same query at six or more different LLMs in parallel, then looks for where they agree, using cosine similarity and a spot of graph centrality to find the consensus answer rather than simply picking whichever model shouted first. The project's own figures claim a hallucination rate under 2%, against 10 to 15% for a single model answering alone. I'll be straight with you: that number is self-reported and there's no published third-party methodology paper sitting behind it, so treat it the way you'd treat any vendor's headline stat, which is to say with one eyebrow raised. What isn't marketing, though, is the architecture producing it. A parliament of models that has to converge before it speaks is structurally less prone to confident nonsense than a single model that has never been asked to check its answer against five colleagues. You don't need to trust the percentage to trust the principle.

 

 

Handing Over the Keys: The Developer Platform

Here is the part I find genuinely satisfying, in the way that only a CIO of thirty years' standing can find an API surface genuinely satisfying: AskDiana ships with a proper Extensions Marketplace, not a roadmap slide promising one.

The design principle, laid down explicitly in the extension guidelines, is that a developer must be able to add full functionality to AskDiana without touching the core application. Extensions live outside the walled garden entirely, running in isolated Docker containers, talking to the platform through an authenticated API (/api/ext, key-authenticated, plus a marketplace and developer-account layer for publishing). A developer applies, gets reviewed in the developer hub, links a GitHub repo for release webhooks, and publishes to an install-from-the-marketplace flow that any tenant can browse.

Rather than trying to out-guess every possible thing a business might want to do with a chat answer, AskDiana treats extensions as post-conversation actions: take this answer and send it to Gamma to become a slide deck, push it to DocuSign, export it to Word. The reference implementation is a working Gamma.app extension, and the Refugee Council's own access widget, RIA_Access, is on the roadmap to graduate from a bespoke integration into a proper first-party marketplace listing that other NGOs running AskDiana could simply install.

This is, not to put too fine a point on it, the correct way to build a platform. Not "here is our chatbot, please enjoy it as delivered," but "here is our chatbot, and here is a documented SDK with example code, so that the businesses using it can build the seventeen strange, specific, industry-particular things we would never have thought to ship ourselves." Sierra Studios didn't wait for us to build a bespoke Galaxy ERP integration from scratch; the platform was already shaped to let that connection exist as a first-class citizen rather than a workaround.

 

Why This Matters More Than the Leaderboards

The industry spends a great deal of energy arguing about which frontier model tops which benchmark this month, as though the winner changes meaningfully what happens when someone at a mid-sized manufacturing firm asks a chatbot why inventory turnover dropped in March. It doesn't, particularly, if that chatbot has never seen the firm's inventory data and is reasoning from vibes. AskDiana's bet, and it's one I happen to think is correct, is that the more defensible competition isn't "which model is cleverest in general," it's "which system actually knows what it's talking about, admits when it doesn't, and lets someone else build the bit it got wrong." Frontier models will keep getting better at being generally clever. AskDiana is getting better at being specifically right, and handing you the tools to make it more right still. That is a considerably harder thing to commoditise.

All Posts