The Abilities of UniDB
A high-performance polyglot storage engine designed to eliminate the boundary between tabular rows, graph edges, and vector similarity.
Three Modalities, One Binary File
Rather than running separate database daemons, UniDB fuses Relational Columnar pages, Compressed Sparse Row (CSR) graphs, and SIMD-aligned vector indices into a single, page-aligned .unidb file layout.
Kernel-Bypass Shared Memory IPC
Producers and consumers communicate directly through a lock-free shared memory ring buffer. With zero TCP socket overhead and zero serialization, ingestion benchmarks exceed 9.1 million ops/sec.
Dual Nanosecond Interval Splitting
Every tuple records both valid time (when an event happened in reality) and recorded time (when the system observed it). Half-open interval splitting guarantees zero overlapping time slices on disk.
Two-Tier LSM CSR Graph
A dynamic frontier absorbs real-time edge mutations before compacting into a static Compressed Sparse Row format. Enables $O(\log N)$ node lookups with instant $O(1)$ slice traversals.
Vector Registers & HNSW
Embeddings are 64-byte row-pitch padded for unaligned, penalty-free AVX-512 and ARM NEON K-Nearest-Neighbor search. Scans vector registers directly without scanning record payloads from disk.
Model Context Protocol Native
Equipped with first-class JSON-RPC 2.0 MCP server capabilities over stdio and Server-Sent Events (SSE). Exposes deterministic contract verification, task tracking, and semantic search directly to LLM agents.
Why Polyglot Convergence Matters
In conventional data architectures, answering a hybrid query—such as "Find all functions semantically similar to this concept, trace their upstream callers 3 hops deep, and filter by commits created in the last 24 hours"—requires orchestrating three distinct databases (Vector DB, Graph DB, and Relational/Time-Series DB).
In UniDB, all three operations happen in the same physical memory space. Graph node pointers resolve directly to columnar offsets, vector indices map directly to graph vertices, and bitemporal filters apply zero-copy without disk record scanning.