RetrievalRouter introduces a joint modality and architecture selection framework for document retrieval, enabling simultaneous optimization of text or multimodal inputs with dense or late-interaction retrieval models. This approach resolves the trade‑off between retrieval accuracy and computational efficiency, allowing scalable deployment in high‑stakes domains such as finance, healthcare, and law. By jointly selecting modality and architecture, RetrievalRouter aims to reduce latency while preserving evidence retrieval quality.