unum-cloud/UStore

What it solves

UStore is a modular, multi-modal transactional database designed for AI and semantic search. It aims to replace the need for multiple specialized databases (like MongoDB for documents, Neo4J for graphs, and Pinecone for vectors) by providing a single ACID-transactional system that can handle various data types with high performance.

How it works

UStore acts as a "build your database" toolkit that defines zero-copy binary interfaces for CRUD operations. It decouples the storage engine from the high-level language drivers, allowing users to swap underlying engines (such as RocksDB, LevelDB, UDisk, or UCSet) while maintaining the same API. It supports multiple data modalities including:

• Blobs: Binary Large Objects, with high-performance options like UDisk for kernel-bypass IO.
• Documents: Support for JSON, MessagePack, and BSON using RFC standards for addressing and updates.
• Graphs: Optimized for large workloads using delta-coding and batch-like processing.
• Vectors: Integrated vector search with support for quantization (8-bit int, 16-bit float) and various distance metrics.

Who it’s for

It is designed for developers building AI applications, semantic search engines, and high-performance data systems who need a transactional database that can store diverse data types (documents, graphs, vectors) without the overhead of managing multiple separate database systems.

Highlights

• Multi-Modal: Single system for blobs, documents, graphs, and vectors.
• ACID Transactions: Guarantees atomicity, consistency, isolation, and durability.
• Modular Engines: Supports multiple backends including RocksDB and the high-performance UDisk (kernel-bypass).
• Broad Language Support: Drivers for Python, C, C++, GoLang, and Java, plus an Apache Arrow Flight RPC interface.
• AI Integration: Comes with Pandas and NetworkX APIs and integrates with USearch and UForm for vector search.