hebbian-robotics/siglip-onnx-server

SigLIP ONNX Server is a small, CPU-only text embedding service built in Rust. It is designed for a specific job: turning search queries into vector embeddings quickly, without needing a GPU. SigLIP 2 is a model that places images and text into the same mathematical space, which lets you compare a text query against a library of image vectors to find matches. The idea behind this project is a split workflow. Building an image search index is heavy work, so that part still runs on GPUs, converting large batches of images or video into vectors ahead of time. But once that index exists, answering a live text search only requires converting one short query into a vector, which is a much lighter task. This server handles just that lighter, live-query side, so a team does not need to reserve expensive GPU capacity just to answer everyday search requests. It exposes an API endpoint compatible with OpenAI's embeddings format, so existing tools and clients built for OpenAI's API can point at this server instead, using their normal integration code with no special wrapper needed. Only two endpoints exist: a health check and the embeddings endpoint itself, keeping the service intentionally minimal. Under the hood, it uses ONNX Runtime to run an exported version of the SigLIP text model efficiently on ordinary CPU hardware, with Rust handling the surrounding server logic like tokenizing text and validating requests. The server requires specific local model files at startup, checks that they match an expected contract, and runs a test inference before it starts accepting real traffic, to catch configuration mistakes early. It can be run directly with Rust's build tools or as a Docker container, and a ready-made container image is published for quick use. This project is best suited for teams already running SigLIP-based image search who need a lightweight, dependable way to serve the text half of that search on CPU.