Vespa.ai Live 2026

Jon Bratseth

All You Need to Solve Everything is to Have the Right Information at the Right Time

In this keynote, Jon Bratseth shares what the next generation of AI infrastructure looks like and why retrieval, ranking, personalization, and real-time data processing must work together in a single system to power modern AI applications. Drawing from decades of experience building large-scale search and recommendation platforms, Jon explores the architectural shifts driving today’s most advanced AI experiences across commerce, search, recommendation, and agentic applications.

Cem Ünsal (Walmart)

Path to Success: How to overcome personal bias, engineering concerns and other pitfalls to take Vespa to production

Cem talk about his personal search journey with Vespa starting in early 2000s. He talks about his approach to making Vespa available for various verticals, emphasizing potential pitfalls and short-term gains. Names of the innocents will be redacted.

This talk includes notes on creating a fast path to production for a low-latency, high-relevance search engine with multi-tenancy in complex domains. It will list pitfalls, key tenets and surprises when delivering fast and at scale, with small teams.

Gregg Donovan

Lessons for State-of-the-Art E-Commerce Search with Vespa

Based on lessons from Etsy’s migration from Solr and multiple other retrieval engines to a single Vespa-based system, this talk explains how model building and evaluation should shape data ingestion. It shows why Vespa should consume model- and evaluation-relevant data from a point-in-time-correct ML data platform instead of relying on separate pipelines that are difficult to reproduce.

The talk also covers three related practices. First, we explore what LLMs need in order to perform migrations from other retrieval engines successfully. Second, we discuss how to distill strong teacher models into fast GBDT rankers that run within the Vespa content nodes. Third, what makes for successful and impactful autoresearch, including the durable execution building blocks that allow optimization pipelines to run for weeks on end.

These lessons form a practical guide for teams building and evolving production e-commerce search with Vespa.

Dainius Jocas (Ravenpack)

Nuances of Binarized Embeddings Based Retrieval

Can we get away without an HNSW index in our billion-scale search system for financial data? This was the starting question in the journey that led to digging deep inside Vespa's internals. 

In this talk I'll focus on the Vespa setup nuances such as:

- Do binarized embeddings need HNSW index?
- How many matches does exact nearest neighbor search expose for the first phase ranking?
- How does the match-phase limiter change the query execution?
- How to combine lexical matching and nearest neighbor search?

Finally, I'll present how far you did we go without HNSW for binarized embeddings based retrieval.

Andreas Eriksen & Marlon Saglia, Vespa.ai

PROBLEM-1 was titled "help": Stories from the Trenches

Marlon and Andreas — two Vespa engineers who have variously broken, debugged, and apologised for Vespa in production — will walk you through what happens when:

- A search company's public search demo gets garbage-collected by S3 because the demo was so stable nothing had changed in 180 days,
- A customer asks Vespa for ten specific document IDs and reliably gets four back (eight engineers, nine hours, the fix is one query parameter),
- A single grouping query with an empty array crashes every search node simultaneously, and
- A customer feeds 25 KB of text containing 11,167 Unicode replacement characters — text so broken our indexer correctly decided it must be binary.

Each story is short, technical, and at our own expense. The through-line: distributed systems where some invariants are local-only, "defensive" code that turned out not to be, optimisations that quietly hide your blast radius — and a healthy reminder that even the people who write search can't always find their own documentation.

Trey Grainger, Searchkernel

Extending Vespa’s Hybrid Search with Wormhole Vectors

In this talk, we introduce “wormhole vectors” and how to use them in Vespa to outperform today's hybrid search approaches. Wormhole vectors are an emerging technique that enables jumping between corresponding regions of disparate vector spaces (sparse lexical, dense semantic, dense behavioral, etc.) at query time to discover relationships that better inform query understanding and ranking.

Christiane Lemke & André Charton, Kleinanzeigen/Adevinta

From Legacy Search to Vespa: What a Real PoC Taught Us

Cem talk about his personal search journey with Vespa starting in early 2000s. He talks about his approach to making Vespa available for various verticals, emphasizing potential pitfalls and short-term gains. Names of the innocents will be redacted.

This talk includes notes on creating a fast path to production for a low-latency, high-relevance search engine with multi-tenancy in complex domains. It will list pitfalls, key tenets and surprises when delivering fast and at scale, with small teams.

Doug Turnbull, SoftwareDoug LLC

Autoresearch for Ranking: Honest Lessons

Can an agent distill better search into executable code? That’s what autoresearch does: tuning algorithms by letting an agent code improvements, measuring, and repeating. But it’s not a silver bullet. It’s another form of machine learning. 

Unlike classic ML, it’s a dance. Like any dance, we guide our partner - nudging agents to push to focus on what’s important. But the partner guides us too! In the cane of an agent with its encyclopedic knowledge.

In this talk I’ll share lessons learned leveraging autoresearch:
- How to build an autoresearcher for your search stack
- Techniques and tools for steering your autoresearcher towards great results
- Where autoresearch might not be the best fit - and you should use a classic ranking model

I’ll share whats worked for me (and not worked). As well as where I see autoresearch headed for the search space.

Alessandro Benedetti & Daniele Antuzi, Sease

Evaluating Search in the LLM Era: Tools for Data & Vectors

This talk explores two open-source tools that leverage large language models to help with search quality evaluation:
the Dataset Generator to generate queries and ratings from your corpus of information and the Vector Search Doctor to diagnose possible causes of the poor performances of your vector search implementation.

Andrius Jokubauskas, Atita Arora, Jon Bratseth, Ravindra Harige

Panel: Building Search for the Modern Age. Moderated by Charlie Hull.

Join leaders from Vespa.ai, Vinted, Searchplex, and beyond as they discuss the future of search. 

Barcamp

A series of short talks submitted on the day and selected by attendee vote.