Vektonn
Overview
Vektonn is a high-performance battle-tested kNN vector search engine for your data science applications. It helps you manage vectors' lifecycle and radically reduces time to market.
See documentation for more info.
Features
- Support for both dense and sparse vectors
- Precise and approximate kNN (AkNN) algorithms
- Scalable architecture that allows to easily handle hundreds-of-GB-worth of vector data
Components
There are three main parts of Vektonn: an API, an Index, and a Data Source.
- The API has methods for search and uploading vector data. It proxies requests to corresponding Indices and Data Sources.
- A Data Source is where all the vectors' data being persistently stored. Currently, a Data Source is implemented using Apache Kafka.
- An Index is an in-memory snapshot of data in Data Source. It updates asynchronously from a corresponding Data Source.
A data from a single Data Source can be spread (sharded) over several Indices to fit in RAM of hosting nodes.
A single Data Source may have several Indices defined on it with different metrics.
Get started
- Read documentation here
- To play with Vektonn locally check out QuickStart guide and example Jupyter notebooks
- To learn how to deploy Vektonn to Kubernetes check out example Helm charts
- If your stack is other than Python or .NET use OpenAPI Specification for Vektonn to develop your own client
Support
If you have any questions or need help with Vektonn please contact us on Slack channel.
Contributing
Repositories
- vektonn is a main repository with implementations for the API, Index shards and Data Sources. It also contains sources for Vektonn's .NET client.
- vektonn-index is a .NET library for finding nearest neighbors in vector space. It provides an implementation of basic functionality for Indices: vector storage and retrieval.
- vektonn-client-python hosts Python client for Vektonn source code.
- vectonn-examples is a repository with various examples of Vektonn usage.
License
Vektonn is licensed under Apache License 2.0.