Switch to ONNX model for default embedding model

[HammadB]

Description of changes

Summarize the changes made by this PR.

• Improvements & Bug fixes
  • None
• New functionality
  • Adds a ONNX port of sentence-transformers all-MiniLM-L6-v2 in order to remove the dependencies on pytorch, sentence-transformers, sentence-piece and other heavy depdencies. This reduced the on disk environment size needed for dependencies to run chroma from ~900MB to ~300MB
  • The ONNX port and verification of its accuracy live in https://github.com/chroma-core/onnx-embedding
  • The ONNX model is hosted on S3 after being generated in the above repo
  • The implementation here runs the model and applies mean-pooling using numpy since that's the final layer.
  • The embedding model will download the model using tqdm to provide the same download experience as before
  • If the model is cached it will not be downloaded.
  • In contrast to before, the model is now ONLY downloaded if used!
  • The net-new dependencies are onnxruntime and tokenizers, both are lightweight.
  • Updated the default model to be this one instead of ST.
  • We create a new DefaultEmbeddingFunction which aliases the ONNX embedding function

Test plan

Added a test to test multiple batches with the new model

Documentation Changes

Are all docstrings for user-facing APIs updated if required? Do we need to make documentation changes in the docs repository? We will need to change the documentation here https://docs.trychroma.com/embeddings#default-sentence-transformers to highlight this fact. This PR will not merge until that documentation is live.

[HammadB]

Chroma-core/onnx-embedding’s “compare_onnx.py” runs two benchmarks to evaluate that the two models are equivalent. Would you prefer I turn it into a test? I didn’t anticipate us ever touching this again so left it hacky.

That should be create_onnx - will fix thanks!

[HammadB]

Docs added - https://github.com/chroma-core/docs/pull/65