Add support for PaLM models, such as chat-bison and text-bison

[ivano-donadi-ennova]

Hi,

I added PaLM models to the available LLMs in guidance. A quick overview can be found in the dedicated notebook in the llm folder.

New features:

• "context" alias for the system role
• "example" role containing input and output blocks to provide demonstration examples, including notebook formatting of examples
• support for library (through VertexAI SDK) and rest calls to google's text genertion and chat models with or without streaming

Issues:

• the select tool requires tokenization of the prompt. However, to the best of my knowledge, there is not an equivalent to tiktoken providing google's models tokenizers. This could still be done by a billable rest call to their tokenization API, but I would like to get community feedback to decide whether to implement this feature or if there are other workarounds.

Tests:

• I tried to replicate all tests done on OpenAI models on PaLM, with the exception of the select tool.

Please let me know what you think!

[xnohat]

Hi,

Google doesn't provide any Tokenizer open-source library, so I think you should change self._tokenizer to a restful api wrapper
original copied from _openai.py

        import tiktoken
        if encoding_name is None:
            encoding_name = tiktoken.encoding_for_model(model).name
        self._tokenizer = tiktoken.get_encoding(encoding_name)

[ivano-donadi-ennova]

Hi,
unfortunately, google's text embedding apis do not seem to include a decoding option, which is required in the 'select' pipeline.