Add multiple embedding providers for SemanticSimilarity

[linear]

Summary

SemanticSimilarity currently defaults to OpenAI's text-embedding-3-small. While extensible via BaseEmbeddingModel, there are no pre-built alternatives. Add support for common embedding providers out of the box.

Motivation

Not all users have OpenAI API keys. Ragas supports multiple providers. Local embedding models (Sentence-Transformers) enable offline testing — important for CI/CD without API key management and for cost-sensitive users.

Implementation Guide

Steps

1. Create pre-built embedding model implementations in libs/giskard-checks/src/giskard/checks/utils/embeddings.py

2. Implement providers:

   • SentenceTransformerEmbedding — uses sentence-transformers library (local, free)
   • Document how to use existing BaseEmbeddingModel for custom providers

3. Add optional dependencies:

   [project.optional-dependencies]
   local-embeddings = ["sentence-transformers>=2.0,<4"]

4. Update set_default_embedding_model() documentation

Example usage

from giskard.checks import SemanticSimilarity, set_default_embedding_model
from giskard.checks.utils.embeddings import SentenceTransformerEmbedding

# Use local embeddings (no API key needed)
set_default_embedding_model(SentenceTransformerEmbedding("all-MiniLM-L6-v2"))

check = SemanticSimilarity(reference="Hello world", threshold=0.8)

Acceptance Criteria

• Sentence-Transformers provider works out of the box
• Clear documentation for custom providers
• Optional dependency — doesn't break core install
• Tests cover: local embeddings produce valid similarity scores

[linear]

OSS-58