smallevals — CPU-fast, GPU-blazing fast offline retrieval evaluation for RAG systems with tiny QA models.

Conceptual path finding in high-dimensional embedding space via Grassmann manifold geometry. Replaces cosine similarity with tangent space distance — paths follow domain-coherent lanes instead of converging to vocabulary hubs. GPU-accelerated Julia compute worker

Small demo project for evaluating RAG retrieval quality using a simple BM25-style retriever and standard metrics (Hit@k, Recall@k, MRR).

📚 Build an offline multimodal RAG system that retrieves grounded answers from diverse document types with full citation support, all in one unified interface.