Vespa
Open-source AI search engine from Yahoo combining vector, text and structured search with machine-learned ranking at scale
Overview
Vespa is a search and serving engine that combines vector search, full-text search and structured filtering with machine-learned ranking, all evaluated close to the data. Rather than only returning nearest neighbors, Vespa lets you write ranking functions and run ML models (including ONNX models and multi-vector approaches like ColBERT) on every candidate, which is why it powers large search, recommendation and RAG systems. It handles real-time updates and scales to billions of documents.
Vespa was built at Yahoo, where it serves many applications, and was spun out as an independent company, Vespa.ai, in 2023 with $31M from Blossom Capital. Users include Spotify, Perplexity, Farfetch, Vinted and Elicit. It is Apache 2.0 open source, with a managed Vespa Cloud.
The Verdict
Who Should Use Vespa?
Best For
- Search and recommendation systems with custom ranking
- Large-scale RAG that mixes keywords, vectors and filters
- Real-time updates on billions of documents
- Teams with search and ML engineering expertise
Not Ideal For
- Simple vector lookup for a prototype
- Small teams that want minimal configuration
- Projects needing a quick MVP
What's Great
- Most flexible ranking of any vector search engine
- Hybrid vector, text and structured search in one query
- Proven at Yahoo and Spotify scale
- Real-time writes with low query latency
- Open source with a managed cloud option
Watch Out For
- Steep learning curve: schemas, ranking profiles and deployment packages
- More operational overhead than focused vector databases
- Smaller community than Elasticsearch or Milvus
- Enterprise pricing requires a sales conversation
Pricing
View all features & details
Key Features
- HNSW vector search with filtering
- BM25 and linguistic text search
- Custom multi-phase ranking with ML models
- Multi-vector and ColBERT-style retrieval
- Real-time updates
- Grouping and aggregation
Use Cases
- Large-scale RAG
- Ecommerce search and recommendations
- Personalized content feeds
- Question answering over large corpora
How It Compares
| Feature | Vespa | Elasticsearch | Milvus | Weaviate |
|---|---|---|---|---|
| Open source | Apache 2.0 | AGPL option | Apache 2.0 | BSD-3 |
| Custom ML ranking | Built in | Learning to Rank | Limited | Limited |
| Hybrid search | Yes | Yes | Yes | Yes |
| Real-time updates | Yes | Near real-time | Yes | Yes |
| Ease of setup | Hard | Moderate | Moderate | Easier |
Is this your tool? Add a badge