Product
KyroDB
A high performance vector database for RAG and AI agents, built around predictable retrieval latency and durable storage.
What it is
KyroDB is a single-node vector database optimized for retrieval-heavy workloads. When agents need context, latency becomes a capability bottleneck, so the system is designed around fast reads and a tiered architecture.
Today KyroDB ships HNSW k-NN search, a two-level Hybrid Semantic Cache (learned frequency prediction plus semantic query matching), a hot tier for recent writes, and durability via WAL and snapshots. The core stack is validated with long-running workload tests and a large automated test suite.
Hybrid Semantic Cache
Two-level L1 cache combining learned hotness prediction and semantic query reuse (73.5% cache hit rate in a 12-hour MS MARCO validation).
Tiered engine
Cache→ hot tier→ cold tier. Recent writes stay fast while the bulk of vectors live in HNSW for efficient k-NN search.
Durability and recovery
Write-ahead log plus snapshots with crash recovery and integrity checks so memory survives restarts and failures.
Validated and hardened
Automated tests covering core correctness, tiering, and recovery behavior.