임베딩 모델 벡터 공간 간 변환을 위한 Python 라이브러리

Make embedding spaces interoperable with simple, drop-in adapters. Bridge embedding spaces. Use adapters, not hacks. all-MiniLM-L6-v2 + Embedding Adapter reaches ~93% of OpenAI’s text-embedding-3-small recall (R@1/5/10) while running locally in just a few ms. Translate embeddings between model spaces locally with a single command. pip install embedding-adapters embedding-adapters embed \ --source sentence-transformers/all-MiniLM-L6-v2 \ --target openai/text-embedding-3-small \ --flavor large \ --text "Where can I get a hamburger?" ^ outputs an embedding and confidence score embedding-adapters is a lightweight Python library and model collection that lets you map embeddings from one model’s space into another’s. Instead of: - Re-embedding an entire corpus every time you change models or providers, or - Locking your search / RAG stack to one vendor’s embeddings, you can: - Embed with a source model (often local / open-source), - Pass those vectors through a pre-trained adapter, and - Use the result in a target embedding space (for example, an OpenAI embedding index). The goal is to make “take vectors from here, make them look like they came from there”: - Easy to adopt – one import, one factory call, one .forward - Consistent – adapters are trained under a known setup (e.g. normalized inputs) - Practical – designed for real retrieval, migration, and experimentation workflows Quality / out-of-distribution (OOD) scoring is supported as an optional diagnostic feature. It can help you understand when an adapter is likely to behave well on your data, but it is not required to start using the library. Real problems this helps with: - Avoid full re-embedding when changing models You already have a corpus embedded with Model A (e.g. a cloud provider). You want to start using Model B (e.g. a locale5 variant) for queries or new content, but re-encoding everything is expensive or disruptive. An adapter lets you map into the existing space instead of re-building the world in one shot. - Local-first or hybrid setups You want to run a strong open-source model locally (for cost, latency, or privacy reasons), while keeping your vector database and relevance logic in terms of a “canonical” target space. Adapters let you keep that target space stable while you change what runs at the edge. - Cross-model interoperability Treat “embedding space” as a contract, not “whatever the current provider happens to be.” Adapters let you plug multiple embedding backends (Hugging Face, OpenAI, etc.) into a shared or slowly evolving space. - Fast experimentation You want to try different source models against a fixed target space / index without rebuilding the entire system every time. Adapters give you a low-friction way to do that. - Extremely Cheap embeddings Run low-cost or local embedding models (MiniLM, e5, etc.) while still operating in a premium target space like OpenAI’s. You keep the retrieval quality of the expensive model for a fraction of the cost, and you only pay the cloud provider when you choose to — not for every embedding. - Fast Local embeddings Local or lightweight models can generate vectors in just a few milliseconds. With an adapter, you keep this speed while still operating inside a stronger target embedding space. This makes retrieval feel instant and dramatically reduces latency for chat, search, ranking, and real-time applications. In short: EmbeddingAdapters turns cross-model compatibility into a first-class, reusable primitive, rather than an ad-hoc alignment script hidden inside a platform or one-off migration projectps! When serving users with familiar or standard questions, waiting ~200ms for a cloud-based embedding model can be unnecessary overhead. By using a local model (or caching strategies) and an adapter layer, you can answer common queries quickly while still aligning with the canonical embedding space. This improves responsiveness and user experience without compromising the integrity of your system - Don't waste time in unnecessary network hops! Intelligent routing for difficult queries When the system recognizes a query as unfamiliar, complex, or requiring higher fidelity, embedding-adapters can help route the request to a stronger or more specialized provider. You maintain a consistent target space while flexibly selecting the best model for this EmbeddingAdapters has the tools for this! Just use our quality endpoints and find out if your query will work, if it won't route to your cloud provider. Mapping between vector spaces is not a new idea in itself. People have aligned word embeddings, distilled models, and trained student/teacher embeddings for years. What is new and different about this project is how that idea is packaged and exposed: - A registry of pre-trained, cross-model adapters you can load with one call, instead of rolling your own alignment for every project. - A focus on model-to-model compatibility, not just query-only tweaks for a single model and corpus. - An explicit design fo