Jo Kristian Bergum

Co-founder of Hornet, a retrieval infrastructure company focused on agentic AI systems. Formerly Chief Scientist at Vespa AI, where he led research on AI-powered search and retrieval, including the native ColBERT embedder and LLM-as-judge work.

Articles in This Vault

• Announcing the Vespa ColBERT Embedder
• Using Approximate Nearest Neighbor Search to Find Similar Products
• E-commerce Search and Recommendation with Vespa
• Improving retrieval with LLM-as-a-judge
• This Is What Agentic Retrieval Looks Like — Hornet; GPT-5 query behavior analysis on BrowseComp-Plus
• Three mistakes when introducing embeddings and vector search — pre-trained models without fine-tuning, single-vector models out-of-domain, and unpriced ANN tradeoffs; source of “an exhaustive search might be all you need”
• Improving Zero-Shot Ranking with Vespa Hybrid Search — part one; BEIR, IR evaluation, and why in-domain scores don’t predict transfer
• Improving Zero-Shot Ranking with Vespa Hybrid Search - part two — tuned BM25 + distilled 22M ColBERT hybrid; 0.481 BEIR average, 12 of 13 datasets
• Improving Search Ranking with Few-Shot Prompting of LLMs — flan-t5-xl generating synthetic training data from three labeled examples; 80.2 nDCG@10 on TREC-COVID

Key Contributions

• Native ColBERT embedder in Vespa with 32x asymmetric compression
• Asymmetric binarization (float queries, int8 documents) for late interaction at scale
• LLM-as-judge retrieval evaluation framework (Vespa)
• Characterization of agentic query workload: long queries, web-search operators, multi-turn compounding (Hornet)
• Zero-shot hybrid ranking on BEIR — tuned BM25 fused with a distilled ColBERT reranker, with distributed min-max normalization computed at the query dispatcher
• Synthetic Query Generation pipeline for domains with no labels: instruction-prompted FLAN-T5, Consistency Filtering, cross-encoder training — released as open notebooks and data

Concepts

• ColBERT
• Late Interaction
• Dense Vector Retrieval
• Hybrid Search
• Agentic Search
• LLM as Judge
• Zero-Shot Retrieval
• Brute-Force Vector Search
• Synthetic Query Generation
• Consistency Filtering
• Score Normalization
• Cross-Encoder

Case Studies

• Vespa - Ranking Without Labels on CORD-19 — the three-post arc from tuned BM25 to a synthetically trained cross-encoder