Turbovec: a Rust vector index with Google Research's TurboQuant algorithm

Turbovec: a vector index in Rust with Google Research's TurboQuant algorithm

The new Turbovec project combines the power of Rust and the TurboQuant algorithm from Google Research to create a next-generation vector index that simplifies the deployment of RAG applications.

TurboQuant Algorithm: Compression Without Training

Google Research developed TurboQuant — an innovative method for quantization (compression) of vector data. This is a key technology for scaling RAG systems and other applications that need to work with large vector databases in memory and on disk.

Traditional compression methods require codebook training — a process where the system analyzes a representative dataset and learns to best compress new vectors. This stage is expensive in terms of time and computational resources: you need to collect data, select hyperparameters, run training, and perform validation.

TurboQuant avoids all of this. The algorithm is based on mathematical analysis of the statistical properties of vectors and can be applied to any data without preliminary training.

The results are impressive: compression reaches 16x. Data that previously occupied gigabytes of memory now fits into megabytes. At the same time, search quality practically does not suffer — the distances between vectors in the compressed space are preserved with high accuracy, ensuring reliable retrieval and accurate search results.

Turbovec: Meeting Rust and Python

Turbovec is an implementation of TurboQuant in Rust with convenient Python bindings. The choice of architecture language is not accidental: Rust provides maximum execution speed without a garbage collector, which is critical for indexes working with billions of vector points.

In such systems, even microsecond delays in search operations can accumulate and lead to significant slowdown of user experience.

The Python interface solves the second problem: it allows machine learning engineers and data engineers to easily integrate Turbovec into their pipelines without rewriting logic in Rust.

This approach is a meeting of two worlds: the performance of a systems language plus the practicality and speed of development in Python.

The architecture assumes the following scenario: the index is created once in Rust for maximum performance, and applications access it through the Python API. This reduces cognitive load on developers and accelerates the development cycle while maintaining maximum efficiency in production.

Application in RAG and Vector Pipelines

The main application of Turbovec is RAG (Retrieval-Augmented Generation) pipelines. In such systems, the principle is simple: take the original text or document from an external source, convert it into a vector using an embedding model, search for relevant documents through vector search, and pass the found results to an LLM for answer generation.

The LLM generates a much more accurate answer based on the context from the found documents than without retrieval.

16x compression provides several practical benefits:

• Lower memory requirements — an index with 1 million 384-dimensional vectors now occupies approximately 250 MB instead of 4 GB
• Faster data transfer — fewer bytes over the network between pipeline components, lower latency in distributed systems
• Cloud storage savings — vector databases are usually billed by volume, so compression directly reduces costs
• Faster search — less data to scan, better caching in processor caches

The absence of a codebook training stage is critical for development speed. Previously, engineers needed to collect a dataset, select hyperparameters, run long compression model training, and debug the results.

Turbovec is ready to use out of the box — deployment in hours instead of days.

What This Means

Turbovec makes high-performance vector search more accessible and simpler to deploy. RAG applications, which previously required expensive cloud infrastructure with large amounts of memory, can now run on modest servers.

This expands accessibility for startups and companies in developing markets that want to control their infrastructure costs and cost-per-query.