DEID-1867 LLM Wrapper

[mrplt-pai]

What's Changed?

An LLM wrapper is added. It takes a string as a prompt that is then deidentified. It then gets sent to the LLM, and the response (alongside the original message) is reidentified. Returns a list of strings as a response. Supports OpenAI, Cohere and VertexAI.

Sample commands to test
llm_response = pai_client.send_redacted_prompt(prompt="My sample name is John Smith", llm_model="openai", model="gpt-3.5-turbo", model_api_key="ADD_MODEL_KEY_HERE")

llm_response = pai_client.send_redacted_prompt(prompt="My sample name is John Smith", llm_model="cohere", model_api_key="ADD_MODEL_KEY_HERE")

llm_response = pai_client.send_redacted_prompt(prompt="My sample name is John Smith", llm_model="vertexai", chat_model_name="chat-bison@001") -- needs to be configured beforehand if I understand correctly.

PR Checklist

• Updated unit tests
• Ran unit tests

[mrplt-pai]

I tested this with Cohere and OpenAI. Working on VertexAI still.
@koryf I'd love your opinion on whether this is what we're looking for. It takes a string as an input, deidentifies it, passes it through the LLM and then reidentifies the incoming message as well as the deidentified prompt.