Add Embeddings created by a Sentence-BERT local model and evaluate how it performs against the OpenAI text-embedding-3-large model.

[JohnFunkCode]

https://pypi.org/project/sentence-transformers/ provides a collection of sentence transformers that can run locally are built to perform well for semantic search use cases.

Ask ChatGPT "How does SBERT compare to using an LLM like OpenAI's text-embedding-3-large mode for embeddings used for semantic search?"

We could create a separate column in the splits table to hold the SBERT embedding, and run automated evaluations to compare how they do with our log of questions.

[JohnFunkCode]

9/7/2025 we built the embedding table and copied all the openAI embeddings to it.

Next Step

• Get retriever and analyzer working from the new embeddings table.
• Add the SBERT embedding into Analyzer. Analyzer needs to check for new docs, and check to see if we need to create embeddings for each split.
• Factor Ebeddingizer out of Analyzer make it look for splits that need don't have a full set of embeddings.
• Make Retriever work with the two embedding models
• Write an evaluator that checks to see if the number of splits returned from each embedding model are different. Expand on this as needed - a diff, and possible AI driven evaluator.