LodeDB
An embedded, multimodal vector database built for local retrieval-augmented generation, positioned as a drop-in, GPU-accelerated backend for frameworks such as LangChain, LlamaIndex, and mem0.
LodeDB is an embedded vector database that runs in-process and stores its index on disk, meant to replace the default in-memory or lightweight stores that LangChain, LlamaIndex, and mem0 use out of the box. It reports a smaller on-disk footprint and lower query latency than those default stores and other embedded options such as sqlite-vec, qdrant, pgvector, lancedb, and chroma across published benchmark comparisons.
It runs the same on-disk index on a GPU when one is available for higher-throughput batched search, in addition to CPU-only operation, and is exact by default rather than relying only on approximate search. A migration path is documented for pointing an existing project's coding assistant at moving its current vector store onto LodeDB.
It is written in Python and released under the Apache-2.0 license.