TLDR: At OpenSearchCon Europe 2026, IBM’s Dom Couldwell argued that truly relevant enterprise search requires design based on user intent (query, journey, and product intent) rather than just the literal keywords typed. While vector search is powerful, going all-in on it introduces significant cost, latency, and compliance-challenging opacity, making a balanced hybrid approach with interpretable lexical search necessary. Ultimately, the success of a search system depends on whether it builds user trust and confidence through explainable and consistent relevance. |
Based on the OpenSearchCon Europe 2026 keynote by Dom Couldwell, Product Manager for OpenSearch, IBM | Watch the full keynote →
Somewhere in the last two years, enterprise search went from being a solved problem to being the most interesting unsolved problem in the industry. The arrival of LLMs made it feel like everything was about to get dramatically better. It also surfaced how many organizations had no idea what their users were actually trying to do when they typed a query.
Dom Couldwell’s keynote at OpenSearchCon Europe 2026 was built around a deceptively simple idea: understanding intent is the key to delivering relevance that actually matters. Not the query—the intent behind it. What decision is this person trying to make? How quickly do they need to make it? How much context is enough? Get those questions right, and you know which techniques to apply. Get them wrong, and it does not matter whether you have the most sophisticated vector search infrastructure in the world.
Three kinds of intent
Couldwell organized his framework around three distinct forms of intent that enterprise search systems need to handle: query intent, journey intent, and product intent.
Query intent is the most familiar, understanding what a user means by the specific words they typed, including ambiguity, synonyms, and the difference between a navigational query and an exploratory one. Journey intent is broader: understanding where the user is in a longer decision-making or task-completion process, and what kind of result will actually help them move forward rather than just answering the surface question. Product intent goes further still, asking what outcome the organization needs from this search interaction, whether that is conversion, resolution, compliance, or something else.
Most enterprise search systems optimize explicitly for query intent and implicitly, imperfectly for the other two. Couldwell’s argument is that you cannot build a search mechanism that truly serves users until all three are part of the design.
The limits of vector-only approaches
One of the most direct parts of the keynote was Couldwell’s challenge to the dominant narrative around vector search. Vectors are genuinely powerful, he acknowledged, they unlock capabilities that were not possible before, and semantic retrieval at scale is a real advance. But organizations that have gone all-in on vector-only approaches are running into real limits: cost, latency, and transparency at scale.
The transparency point is particularly important and often underweighted. Lexical search such as BM25, inverted index, keyword matching, is interpretable. You can explain to a stakeholder, a compliance officer, or an auditor why a particular result appeared. Vector search, by contrast, produces results through similarity in high-dimensional space that is difficult to explain in human terms. For regulated industries, that opacity is not just uncomfortable; it can be a compliance problem.
Couldwell was direct about where he thinks the OpenSearch community can play a constructive role: as a voice of reason in a space where the “just add vectors” narrative has been overapplied. The community is well-positioned to help practitioners understand where the balance between lexical and semantic retrieval actually lies, rather than chasing the next layer of complexity.
The goal Is confidence, not just answers
The closing argument of the keynote was the one that stuck: the goal is humans and AI acting faster with confidence. Search that returns results does not accomplish that goal if users do not trust those results. Trust comes from relevance, from explainability, from consistency, from a system that behaves in ways users can develop an accurate mental model of.
That framing shifts the evaluation criteria for enterprise search investment. It is not enough to ask whether your system returns semantically relevant results. The question is whether users trust it enough to act on what it gives them. If they do not, the underlying technology does not matter.
Couldwell delivered this as a practical call to the OpenSearch community rather than an abstract observation. The techniques and tools exist to build search systems that meet this bar. The OpenSearch ecosystem—hybrid retrieval, neural ranking, transparency tooling, the open governance model that allows organizations to inspect and trust what they are running—is well positioned to deliver on it. The work is in applying them with the right understanding of intent.
