Jo Kristian Bergum

Jo Kristian Bergum works on search and neural ranking, including combining keyword scores with token-level representations and retrieving long documents for retrieval-augmented generation (RAG). He is CEO of Hornet, which builds search for agents, and formerly served as Vespa’s chief scientist. His published evaluations emphasize testing neural methods against strong keyword-search baselines.[3][1][4][5]

In an essay bearing a January 3, 2025 publication date, Bergum argues that vector search belongs within a broader retrieval toolbox: “What we label as "vector databases" are, in reality, search engines with vector capabilities.” He welcomes the accessibility of pretrained embeddings but argues that practical retrieval also requires text matching, filtering, relevance tuning, and multiple ranking strategies. In his account, adding vector storage and similarity search to a database does not by itself make it a complete retrieval engine.[6]

Earlier, as Vespa’s chief scientist, Bergum published evaluations that give concrete technical context to this position: a 2023 comparison of keyword, neural, and hybrid ranking, followed by a 2024 announcement extending ColBERT scoring to long documents.[4][5]

In his January 2023 zero-shot ranking article, Bergum described the Vespa team’s approach to searching without task-specific training data. They established a BM25 keyword baseline by combining title and text scores, then used a distilled MiniLM-based ColBERT model to rerank candidates. Their hybrid method took the top 2,000 BM25 results, scaled both BM25 and ColBERT scores to a 0–1 range, and combined them linearly. A stateless custom searcher performed normalization across the returned candidates, addressing the fact that individual content nodes did not know the global score extremes.[4]

Bergum reported that the hybrid ranked best among the three evaluated methods on 12 of 13 out-of-domain BEIR datasets, raising average nDCG@10 from 0.453 for BM25 to 0.481. The evaluation omitted four unavailable datasets and CQADupStack, and the average excluded the in-domain MS MARCO dataset. The exceptions mattered: BM25 beat the hybrid on CLIMATE-FEVER, while ColBERT alone beat it on MS MARCO. Bergum argued that weak BM25 baselines can exaggerate apparent neural-ranking progress, particularly when testing generalization beyond a model’s training domain.[4]

In March 2024, Bergum announced Vespa’s long-context ColBERT implementation. ColBERT retains contextualized vectors for individual tokens rather than pooling an entire text into one vector. The extension stored representations from multiple encoder windows within one retrievable document, preserving access to its full text and metadata. Encoder windowing remained necessary, and tokens in different windows did not share encoder attention; the change concerned how their representations were stored and scored.[5]

Bergum explained two scoring options. Context-level MaxSim scored each window separately and used the strongest window as the document score. Cross-context MaxSim let each query token’s strongest match anywhere across the document’s windows contribute, potentially helping when relevant information was spread across several sections. Both methods used BM25 to shortlist documents before neural reranking.[5]

On the MLDR English split—200,000 documents and 800 test queries—Bergum reported that both methods outperformed the BM25 baseline, with context-level scoring strongest. The experiment reranked 400 candidates using 1,536-character windows with no overlap. He cautioned that this was one dataset, that its construction could affect the result, and that the better scoring method would depend on the search task. His stated commitment was explicit: “We will not hide simpler baselines to sell storage and compute units without significant benefits for the application.”[5]