Awesome Search KG

Vespa

4 min read

Vespa

What They Build

Vespa is an open-source big data serving engine originally developed by Yahoo and now maintained as an independent open-source project (vespa.ai). Unlike Elasticsearch (document-centric) Vespa was built with ML-native ranking in mind from the start — it supports tensor computations, native embedding models, and multi-phase ranking as first-class features.

Vespa is particularly strong for use cases that mix retrieval, ranking, and ML inference in a single platform without an external reranking tier.

Search Contributions

Models and Embeddings

  • ColBERT embedder — native integration of ColBERT late-interaction model into Vespa with asymmetric binarization (32x compression with minimal quality loss). Led by Jo Kristian Bergum. See Late Interaction in Vespa for the indexing/scoring/scaling mechanics
  • Native HNSW approximate nearest neighbor for dense vector retrieval
  • BM25 + dense vector Hybrid Search as a built-in feature

Methodology

  • Zero-shot ranking without labels — a three-post programme establishing that a tuned BM25 baseline fused with a small distilled ColBERT reranker beats either component on BEIR (0.481 vs 0.453 avg nDCG@10), then that Synthetic Query Generation with a 3B open LLM closes the remaining gap on a specific domain. See Vespa - Ranking Without Labels on CORD-19
  • LLM-as-Judge for retrieval evaluation — demonstrated how to use LLMs to generate relevance labels and compute NDCG; showed strong correlation with human judgments (Spearman ρ ≈ 0.85–0.90). Early and influential demonstration of the LLM as Judge pattern for retrieval

Architecture

  • Multi-phase ranking: cheap first-stage rankers → expensive ML models only on candidates. Native LTR via GBDT (XGBoost/LightGBM) and ONNX models in ranking expressions — see Vespa Learning to Rank
  • Tensor computations at serving time without external ML serving infrastructure
  • Native support for structured filtering alongside dense vector search
  • Distributed exact nearest-neighbour search — an exhaustive scan can be spread across nodes in parallel to hold latency down, and restricted to a subset by query-engine filtering, so brute force stays viable further up the corpus-size curve than it would in a single process

Position in the Ecosystem

Vespa is a direct alternative to Elasticsearch + separate ML serving tier. The value proposition: run BM25, dense vector ANN, and ML re-ranking in a single engine. The tradeoff: steeper learning curve, smaller community than Elasticsearch.

People

  • Jo Kristian Bergum — former Chief Scientist, Vespa AI (now co-founder of Hornet); author of the ColBERT embedder and LLM-as-judge work
  • Marianne Haugvaldstad — Developer Intern; HNSW exploration and tooling
  • Brage Vik — Developer Intern; HNSW exploration and tooling

Articles

Concepts

ColBERT · Late Interaction · Late Interaction in Vespa · Learning to Rank · Vespa Learning to Rank · LLM as Judge · Hybrid Search · Dense Vector Retrieval · Reranking · Brute-Force Vector Search · Zero-Shot Retrieval · Score Normalization · Synthetic Query Generation · Consistency Filtering · Cross-Encoder

Case Studies

Datasets

  • BEIR · TREC-COVID — the benchmarks the zero-shot ranking work is measured on

Graph View

Backlinks