Thoughts on dynamic LoRA adapter support for embedding models in TEI?

### Feature request Text Embeddings Inference (TEI) currently focuses on serving static embedding models efficiently. However, there is no apparent support for loading LoRA / PEFT adapters or dynamically selecting or swapping adapters at runtime. I wanted to ask whether dynamic LoRA adapter support for embedding models is something the maintainers have considered or see as aligned with TEI’s long-term direction. ### Motivation Many embedding-centric pipelines (e.g., RAG, multi-domain retrieval, task-specific similarity search) rely on PEFT / LoRA fine-tuning to adapt a shared base model to multiple downstream tasks. At the moment, TEI appears optimized for serving static embedding models, which means users who rely on LoRA adapters typically have to either: - Spin up separate TEI instances per adapter, or - Merge LoRA weights into full models ahead of time This raises operational overhead and limits flexibility compared to modern inference setups where adapters can be selected dynamically. The motivation behind this issue is to understand whether dynamic LoRA adapter support for embedding models is something the maintainers have considered or see as aligned with TEI’s long-term direction, or whether TEI is intentionally scoped to static models by design. ### Your contribution At this stage, I’m primarily looking for guidance on project direction and maintainers’ intent. If dynamic LoRA support is considered in-scope for TEI, I’d be happy to: - Help with design discussion or an RFC-style proposal - Prototype or contribute a PR aligned with existing TEI architecture and constraints I’m also happy to adapt to whatever contribution process the maintainers prefer.