valentinus
A thread-safe vector database for model inference embeddings, built in Rust on top of LMDB.
valentinus is a thread-safe vector database intended for model inference, storing its data inside LMDB rather than a separate database server. It uses lmdb-rs for the underlying database bindings, ndarray for numpy-style array handling, wincode and serde for serialization, and ONNX Runtime for computing embeddings.
Configuration is controlled through environment variables covering the working directory for the database, the maximum LMDB map size, the number of parallel ONNX execution threads, custom embedding dimensions for non-default models, and which named environment (such as test or prod) to use. Its test suite requires the all-MiniLM-L6-v2 ONNX model and tokenizer files, obtained from Hugging Face or built directly.
The project describes itself as inspired by a ChromaDB Python tutorial, and ships example code alongside its test suite.