add embeddings to TestsetGenerator

[hunter-walden2113]

Addresses the "no embeddings found" and "API Connection error" issues.
Specifically issues: 1546, 1526, 1512, 1496

Users have reported that they cannot generate a Testset because they get API connection errors, or their knowledge graph does not have the embeddings. This is due to the use of the default LLMs and Embedding models via llm_factory and embedding_factory. The errors are occuring becuase the users do not have OpenAI credentials in their environment because they are using different models in their workflow.

Issue to solve is to prevent the default_transforms function from using the llm_factory by forcing the user to add both an embedding model and llm model when instantiating TestsetGenerator.

1. Added embedding_model as an attribute to TestsetGenerator.
2. Added embedding_model: LangchainEmbeddings as a parameter to TestsetGenerator.from_langchain
3. Changed the return from TestsetGenerator.from_langchain to return cls(LangchainLLMWrapper(llm), LangchainEmbeddingsWrapper(embedding_model), knowledge_graph)
4. Added both an llm and embedding_model parameter to TestsetGenerator.generate_with_langchain_docs

[jjmachan]

everything looks good but just ones point here. The TestsetGenerator object does not need an embedding model since it does not use it why synthesizing the testset (embeddings are only need for transforms).

I love make the llm and embedding model and forcing the user there - will help solve a lot of the issue

Thanks a lot for making this contribution 🙂 ❤️, let me know what you think about the point I made

[hunter-walden2113]

Ok thanks for the quick review. The reason I added embedding_model as an attribute is because in line 61, I made it so you can pass your embedding model to the from_langchain method which returns an instance of TestsetGenerator, containing both the LLM and embedding model (line 68). It allows the TestsetGenerator object to have an embedding model attributed to it, so if the user does not pass an embedding model to the generate_with_langchain_docs for the transforms_embedding parameter, it can use the embedding model that is attributed to the class, and the method will not throw an error (see line 108).

I do understand your point though. And we can remove that if you want. Thanks!

[jjmachan]

I think it is okay for now - merging this in 🙂