Add save/load to Embeddings

[TomeHirata]

Closes #8807

[Copilot]

Pull Request Overview

Adds save/load functionality to the Embeddings class, enabling persistence of embeddings indices to disk for fast loading without recomputing embeddings.

• Implements save(), load(), and from_saved() methods for the Embeddings class
• Adds comprehensive test coverage for save/load functionality including error handling
• Removes the TODO comment about adding save/load methods

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
dspy/retrievers/embeddings.py	Implements save/load methods with pickle for config, numpy for embeddings, and FAISS index persistence
tests/retrievers/test_embeddings.py	Adds comprehensive tests for save/load functionality and error cases

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[chenmoneygithub]

do we need to reinitialize the search_fn?

[TomeHirata]

Yup, this is because from_saved bypasses __init__

[chenmoneygithub]

How do we want this load to be called? Do we want users to first create an Embedding instance then call embedding.load(), or make it a class method that return a loaded embedding?

[TomeHirata]

This is actually only for consistency with other APIs like module.load. I guess mostly people will use from_saved. Do you think we should make load a classmethod?

[chenmoneygithub]

Gotcha, I am asking because of the line self.search_fn = Unbatchify(self._batch_forward), if we do:

embedder = dspy.Embeddings(...)
embedder.load(...)

do we still need this self.search_fn = Unbatchify(self._batch_forward)?

[chenmoneygithub]

LGTM after #8818 (comment) is resolved, thank you!