Universities don’t compete on content—they compete on clarity. See how we shipped a trustworthy AI help portal for the University of Helsinki in 3 months, turning scattered Drupal ecosystems into sourced answers.
To get the most from this session, attendees should have:
- A basic understanding of Drupal concepts (content types, taxonomy, permissions, workflows)
- Familiarity with site search fundamentals (indexing, relevance, analytics)
- General awareness of what LLMs and RAG (retrieval-augmented generation) are (no deep AI background required)
- Experience working with content-heavy organizations (universities/public sector/enterprise) is helpful but not mandatory
Wunder builds web services for multiple education-sector organizations in Finland. In this session we present a production case study: HelsinkiUni Help, a self-service portal for the University of Helsinki that provides a single search experience across multiple university sources and produces AI-generated answers grounded in cited content.
We’ll walk through the key decisions, trade-offs, and delivery tactics that enabled a fast timeline (~3 months) without sacrificing trustworthiness:
- The problem in the education vertical: fragmented knowledge, multiple stakeholders, high expectations for accuracy and accessibility
- Solution approach: RAG-based semantic search across multiple sources, paired with answer synthesis and conventional results
- Trust and safety design: guardrails, “don’t answer when uncertain,” source visibility, and fallbacks
- Quality engineering: relevance thresholds, adaptive filtering, evaluation and regression testing mindset
- Operational excellence: automated content pipelines (crawl → process → embed → index), performance targets, and observability
- Continuous improvement: search analytics and user feedback loops that reveal content gaps and service friction
- What we’d do differently next time: governance, metadata strategy, feature flags/A-B rollout, multilingual and persona-aware retrieval
After the session, participants will be able to:
- Explain when RAG is the right approach for institutional self-service (and when it isn’t)
- Design a trustworthy AI answer UX (sources, fallbacks, refusal behavior) suitable for education/public-sector contexts
- Apply practical quality guardrails for retrieval and generation (thresholds, cutoffs, “safe defaults”)
- Set up evaluation and observability practices to prevent regressions and diagnose failure modes
- Build an improvement loop using search analytics + user feedback to prioritize content fixes and product iterations
- Take away a reusable delivery playbook for shipping an AI-assisted search/answer capability under real schedule constraints
