From Search Chaos to AI – How to Optimize RAG in Knowledge-Intensive Domains en no

Language models can produce impressive responses – but also hallucinations. In a project for Landinfo, the Norwegian Country of Origin Information Centre for the immigration authorities, we explored how state-of-the-art Retrieval-Augmented Generation (RAG) can be used to extract relevant, verifiable information quickly and accurately – without compromising on quality and trust.

By applying a sequential decision-making process for information retrieval – inspired by the Markov Decision Process (MDP) – we treat each search as a step in a continuous improvement loop. This provides a framework for optimizing each part of the retrieval pipeline, allowing better control over generation and dynamic adaptation of responses based on the available information.

In this session, you'll get insight into how Landinfo uses RAG and sequential decision-making to optimize information retrieval and knowledge access in demanding domains. We’ll demonstrate how the technology can be practically implemented to ensure precision and traceability.

This presentation is suitable for developers, technologists, and anyone curious about how generative AI can be applied in knowledge-intensive environments with high demands for traceability and precision. No prior knowledge is required.

Fra søkekaos til AI-kurator - Hvordan optimalisere RAG i kunnskapsintensive domener en no

Språkmodeller kan gi imponerende svar – men også hallusinasjoner. I et prosjekt for Landinfo, utlendingsforvaltningens fagenhet for landinformasjon, har vi utforsket hvordan state-of-the-art Retrieval-Augmented Generation (RAG) kan brukes til å hente ut relevant, etterprøvbar informasjon raskt og presist – uten å gå på kompromiss med kvalitet og tillit.

Ved å bruke en sekvensiell beslutningsprosess for informasjonsinnhenting – inspirert av Markov Decision Process (MDP) – kan vi behandle hvert søk som et steg i en kontinuerlig forbedringsprosess. Dette gir et rammeverk for å optimalisere hvert ledd i informasjonsinnhentingen, sikre bedre kontroll over genereringen og tilpasse svarene dynamisk etter tilgjengelig informasjon.

Her får du innsikt i hvordan Landinfo bruker RAG og sekvensiell beslutningstaking for å optimalisere informasjonsinnhenting og kunnskapstilgang i krevende domener. Vi viser hvordan teknologien kan implementeres i praksis for å sikre presisjon og sporbarhet.

Presentasjonen passer for utviklere, teknologer og andre som er nysgjerrige på hvordan generativ AI kan brukes i kunnskapsintensive miljøer som stiller store krav til sporbarhet og presisjon. Du trenger ingen forkunnskaper.

When is a Model Good Enough? A Year of Speech-to-Text at UDI en

When is a speech-to-text model actually good enough?
Norwegian asylum interviews are among the hardest speech-to-text scenarios a government can face. Three participants. Two languages switching mid-sentence. An interpreter whose code-switching breaks most ASR models. A legally binding record where a mistranscription can shape the outcome of a protection claim. And a data regime where the vendors you most want to test are the ones you are least allowed to test on real data.

Over the past year at UDI, we ran a pilot to answer exactly that question. We evaluated nine systems on theater interviews and historical audio, iterated through four architectures on Azure, from GPU VMs running our own models to an external transcription API, ran seven live asylum interviews with the one model our privacy approvals allowed, and wrote our own speaker separation when off-the-shelf tools failed. The honest answer at the end of the pilot: not yet, but closer than before.

This talk walks through what we learned, both technically and about the three people in the room. What worked in the models and what plateaued. What the live interviews revealed that no benchmark could. And how real-time transcription changes the work of the caseworker, the interpreter, and the person being interviewed. If you are building STT for Norwegian, or evaluating AI systems in regulated domains, this session shows you what that work actually looks like from the inside of a real government project