At a glance
| Product type | Open source search, ranking and vector engine, plus a managed cloud |
| Vector index types | HNSW (approximate) or exact search on tensor fields |
| Hybrid search | nearestNeighbor combined with text matching in YQL and ranking expressions |
| Metadata filtering | YQL filters combined with nearestNeighbor |
| Embedding integration | Built-in embedders: Hugging Face (ONNX), ColBERT, SPLADE locally; VoyageAI, OpenAI and Mistral APIs |
| Scaling and multitenancy | Container and content clusters scaled by allocated resources |
| Deployment (managed or self-hosted) | Self-hosted open source, Vespa Cloud, Vespa Cloud Enclave, Enterprise self-hosted licence |
| Licensing and pricing model | Apache 2.0 engine; Vespa Cloud per vCPU, memory, disk and GPU hour |
Vespa fits teams building large search, recommendation or RAG systems that need vector search, text relevance and custom ranking models evaluated together. It has a steeper learning curve than a simple vector database, but transparent per-resource cloud pricing and an Apache 2.0 engine you can run yourself.
Overview
Vespa began as a search engine in 1997, became part of Yahoo in 2003 and was open sourced in 2017. In October 2023 Yahoo spun it out as Vespa.ai, an independent company. The engine stores documents with fields, including tensor fields for vectors, and evaluates queries and ranking expressions on the content nodes. The GitHub repository says a new release is made from the master branch most weekday mornings, so there is no single long-lived version to track.
Vector search uses the nearestNeighbor query operator, which requires a targetHits value and feeds the closeness() rank feature. Search is exact unless the field has an HNSW index, in which case it is approximate. Because the operator is part of Vespa's YQL query language, it can be combined with text matching and filters in one query to build hybrid retrieval, and a mixed tensor can hold several vectors per document.
Key features
- nearestNeighbor operator for exact or HNSW-based approximate search, with six documented distance metrics set per field
- Multiple vectors per document through mixed tensors, using the closest vector per document
- Hybrid retrieval by combining nearestNeighbor with text matching and filters in YQL
- Built-in embedders that create vectors at feed and query time with the embed function (Hugging Face ONNX models, ColBERT, SPLADE, or VoyageAI, OpenAI and Mistral APIs)
- Binary embeddings with Hamming distance computed using xor and pop count instructions
- Vespa Cloud managed service and Vespa Cloud Enclave, which runs in your own AWS, Azure or GCP account
Integrations and ecosystem
Pros and cons
Pros
- Vector search, text relevance and custom ranking in one engine
- Published per-resource cloud prices with a 300 USD trial credit
- Apache 2.0 engine that can be self-hosted
- Enclave option keeps data in your own AWS, Azure or GCP account
Cons
- More concepts to learn (schemas, tensors, YQL and rank features) than a simple vector API
- Self-hosted open source builds come with community support only and no SLA, per the pricing page
- Premium cloud support has a 20,000 USD per month minimum
- Enclave users pay their cloud provider's costs on top of Vespa Cloud costs
Pricing
The open source engine is free. Vespa Cloud is billed per hour for allocated resources; on 7 November 2026 the pricing page listed, in USD, Basic at 0.10 per vCPU-hour, 0.01 per memory GB-hour, 0.0004 per disk GB-hour and 0.07 per GPU GB-hour; Professional at 0.145, 0.0145, 0.0005 and 0.10; and a Startup plan at 0.05, 0.005, 0.0002 and 0.03. Premium is custom-priced with a 20,000 USD per month minimum. Only container and content clusters are charged, unit prices fall with volume, and a 300 USD trial credit needs no credit card. Enterprise self-hosted licensing is annual and by quote.
Pricing checked on the vendor's official page on 7 November 2026. Prices change; confirm before buying.
Who it is best for
Engineering teams building large-scale search, recommendation or RAG systems that need hybrid retrieval and custom ranking evaluated together.
Typical use cases
- E-commerce search combining keyword relevance, vectors and business ranking signals
- Content and product recommendations
- RAG retrieval with multiple vectors per document (for example per chunk)
- Personalised news or feed ranking
Who should look elsewhere
- Small projects that only need to store and query embeddings, where pgvector or a lightweight vector database is quicker to adopt.
Verdict
Vespa is a strong engine for teams whose search quality depends on combining vectors, text and machine-learned ranking at scale, and its cloud prices are unusually transparent. For simple similarity search it is more engine than most teams need.
Frequently asked questions
Is Vespa open source?
Yes. The Vespa engine is licensed under Apache 2.0 on GitHub. Vespa Cloud, Premium support and Enterprise self-hosted licences are commercial offerings from Vespa.ai.
How much does Vespa Cloud cost?
On 7 November 2026 the Basic plan listed 0.10 USD per vCPU-hour, 0.01 USD per memory GB-hour and 0.0004 USD per disk GB-hour, with higher Professional rates and a cheaper Startup plan. New users get 300 USD of trial credit without a credit card.
Is Vespa still part of Yahoo?
No. Yahoo spun Vespa out as an independent company in October 2023. Yahoo kept a stake and a board seat and remains a major user.
Does Vespa support hybrid search?
Yes. The nearestNeighbor operator is part of the YQL query language, so it can be combined with text matching and filters in one query, with ranking expressions deciding the final order.
Sources
- Vespa pricing
- Yahoo spins out Vespa (press release)
- Vespa nearest neighbor search
- Vespa Cloud Enclave
- Vespa on GitHub
Checked November 2026.
Listed in: Vector Databases & AI Search.
How we research reviews and what star ratings mean: our editorial method.