Tldr we built the fastest and most compact embedded vector database in the world (github.com)

Egoist Machines, Inc. has unveiled LodeDB, claimed to be the fastest and most compact embedded vector database globally, designed to enhance local retrieval-augmented generation (RAG) systems. LodeDB distinguishes itself with an on-disk footprint significantly smaller than that of competitors, while also operating at unprecedented speeds—recording sub-millisecond write times and GPU-accelerated search capabilities that vastly outperform existing solutions. For instance, it demonstrates batched retrieval performance up to 50,326 queries per second on high-end GPU hardware, and durability is ensured through a novel approach that commits only changed rows, allowing for fast operations even at large scales. The significance of LodeDB for the AI/ML community lies in its efficiency and versatility as a memory backend, compatible with popular frameworks like LangChain, LlamaIndex, and mem0. With features such as a compact storage system utilizing 2/4-bit vector packing and a hybrid search mode that combines vector and lexical matches, LodeDB positions itself as a robust choice for applications requiring fast and reliable data retrieval. Furthermore, its open-source Apache-2.0 licensing encourages broader adoption and contribution from the community. Overall, LodeDB addresses critical performance constraints faced by existing vector databases, especially in embedded contexts, making it a compelling option for developers aiming to streamline AI system operations.