PROVIDER Β· Unified search and recommendation engine
Vespa
Vespa is Yahoo's open source large-scale search and recommendation engine combining keyword and vector recall with ML fine ranking and real-time features in one system, ideal where ranking complexity would otherwise need several systems.
Pros
Cons
Product Information
Detailed Notes
Overview
Vespa was built internally at Yahoo from 2003 and open sourced in 2017 as the engine behind Yahoo's search and recommendations, now commercially supported by Vespa AI, distinguished by combining retrieval, machine-learned ranking, and real-time feature computation in one system.
For Backend & Database work the value is ranking as a model, since beyond keyword and vector recall it executes full ranking expressions at query time including tensor operations and ONNX inference, collapsing recall, coarse ranking, and fine ranking into one engine to avoid multi-system latency and consistency costs.
The barrier is learning curve and operations, because application packages define schemas, ranking profiles, and deployment with concepts far from Elasticsearch, cluster operations need Java and distributed experience, and the community is smaller than Elasticsearch with little Chinese material.
Core Strengths
- Retrieval, ML ranking, and real-time features in one system reduce multi-system stitching
- Native hybrid retrieval fuses vectors and keywords during ranking
- Ranking expressions run tensors and ONNX inference without external services
- Single clusters hold billions of documents at millisecond latency, proven at Yahoo
Product Ecosystem
Application packages and schemas
Declare document structure, indexing, and ranking configuration for atomic deployment.
Ranking and features
Rank profiles define multi-phase ranking with tensor computation, ONNX models, and live field updates.
Deployment
Self-managed clusters and Vespa Cloud billed by node and resource size.
Drawbacks
- Concepts differ sharply from Elasticsearch with a steep learning curve
- Cluster operations need Java and distributed systems experience
- Smaller ecosystem than Elasticsearch with little Chinese material
- Powerful but complex configuration, overkill for simple site search
Use Cases
- Unified large-scale search and recommendation (β β β β β ) - Recall and ranking together
- Machine-learned fine ranking (β β β β β ) - ONNX and tensor ranking
- Real-time personalised recommendations (β β β β ) - Live feature updates
- Simple site search (β β ) - Overweight, Elasticsearch fits better
Pricing
| Plan | Price | Scope |
|---|---|---|
| Vespa open source | $0 | Apache 2.0, self-managed |
| Vespa Cloud | Per node and resource-hour | Managed with elastic scaling |
| Enterprise support | Contract quoted | Architecture consulting and SLAs |
The open source edition is free while clusters demand substantial memory and CPU, and where ranking logic is complex and would otherwise need several systems, unification can lower total cost, with managed per-node billing fitting teams strong on ranking but short on operations.
FAQ
-
Vespa or Elasticsearch? Vespa for ML fine ranking, real-time features, and recommendation fusion, Elasticsearch or OpenSearch for a mature ecosystem, easy onboarding, and log search.
-
Can it do vector-only retrieval? Yes, though ranking strengths go unused and dedicated stores like Qdrant or Milvus stay lighter for pure vector search.
-
How are ranking models deployed? Export to ONNX and reference it in rank profiles, with features read directly from document fields and query parameters without a separate feature service.