Martin Bayton
Key Takeaways from the BCS annual conference on information retrieval
Last week I attended Search Solutions 2025, the British Computer Society’s annual gathering of search and information-retrieval specialists. As always, it brought together researchers, practitioners and technologists to explore where search is heading – especially as AI continues to reshape expectations.
Here’s a brief wrap-up of the themes and insights that stood out on the day.
Search Evaluation Takes Centre Stage
Several talks focused on the growing challenge of evaluating search systems in an AI-first world. Alessandro Benedetti showcased his LLM-powered Retrieval, Reasoning and Extraction (RRE) dataset generator, which automates the creation of queries and relevance judgements – traditionally slow and costly work for search teams. Gianna Cipponeri built on this theme with a framework for evaluating citation quality and context faithfulness in RAG systems, highlighting the importance of grounding AI answers in trustworthy source material.
The User Still Matters
A strong thread throughout the day was the importance of supporting human intelligence, not replacing it. Orland Hoeber argued for search interfaces that help people explore, analyse and reason – especially in academic contexts where simply generating an answer isn’t enough. Mark Harwood expanded on this with his ‘plasticine’ embeddings, showing how interactive clustering makes vector spaces more interpretable and lets structure emerge naturally from the data. Research on people with Aphasia offered a powerful reminder that today’s search tools still disproportionately rely on language skills and that inclusive design requires much more attention.
Rethinking How We Retrieve
Sean MacAvaney challenged long-standing assumptions around re-ranking, warning that the classic cascading approach can discard relevant documents too early. His alternative – pulling in additional candidates during re-ranking – demonstrated how rethinking fundamentals can unlock meaningful gains in recall and effectiveness.
AI in Professional Search: Clinical, Legal and Enterprise
The afternoon highlighted the realities of deploying AI in high-stakes domains. Adam Roegiest (Zuva) discussed the challenges of evaluating IR systems in legal tech, where the cost of an error is extremely high. Cedric Ulmer (France Labs) shared how Datafari is integrating GenAI features – securely and transparently – into an open-source enterprise search offering. And in the clinical world, Jon Brassey’s presentation on the evolution of the Trip medical database showed how AI can genuinely improve decision support. Through tools like AskTrip, clinicians can receive rapid, referenced, evidence-based answers – keeping transparency, provenance and user control at the heart of the experience. It was one of the clearest examples of responsible AI being applied to solve a real and urgent user problem.
My Key Takeaways
- Evaluation is more important than ever
LLMs and RAG require new ways to measure trust, grounding and relevance. - Human-centric design wins
Interpretable systems and inclusive design matter as much as model performance. - AI brings opportunity—but also complexity
Hallucinations, content quality and LLM usage costs all remain major challenges. - Real-world deployments are catching up with research
From clinical and legal use cases to enterprise search, we’re seeing more responsible, practical AI implementations.
Have More Questions? Contact Us!
I hope you’ve found this summary of the event useful. The attendees were definitely among the best in the search industry. If you have any questions, please free reach out and Contact Us to schedule a free consultation.
– Martin