Add RAG example using mlx-lm hidden state embeddings

[ManjushaMotamarry]

Summary

Adds a minimal Retrieval-Augmented Generation (RAG) example to mlx_lm/examples/.

What this example demonstrates

• How to generate sentence embeddings using mlx-lm hidden states (mean pooling across the last transformer layer)
• How to compute cosine similarity between a question and a set of documents
• How to retrieve the most relevant document and inject it into the prompt as context
• How to generate a grounded answer using mlx-lm

Why no external vector DB

Intentionally dependency-free — uses numpy for cosine similarity and mlx-lm itself for embeddings, keeping everything within the MLX ecosystem.

Files changed

• mlx_lm/examples/rag.py — the RAG example
• tests/test_rag.py — unit tests for retrieve() and cosine_similarity()
• README.md — added link to the new example

Tests

All 5 tests pass:

• test_identical_vectors
• test_orthogonal_vectors
• test_opposite_vectors
• test_retrieves_most_similar_document
• test_single_document_always_returned