HISTORIC STRUCTURE IN WATSONX.AI FOUNDATION MODELS CONSUME

Hello,

In the IBM Watson AI foundation models' APIs, there is currently no standardized method to submit historical data (several previous interactions with the LLM) so that the model can understand that the conversation pertains to a specific context provided in earlier interactions. For instance, in ChatGPT, if a user first states their name as Isa and then asks if the AI remembers their name, the interaction might look like this:

{'role':'system', 'content':'You are friendly chatbot.'},
{'role':'user', 'content':'Hi, my name is Isa'},
{'role':'assistant', 'content': "Hi Isa! It's nice to meet you. \Is there anything I can help you with today?"},
{'role':'user', 'content':'Yes, you can remind me, What is my name?'}
{'role':'assistant','Your name is Isa. How can I further assist you today, Isa?'}

For the AI model to respond accurately, we need to send a set of previous interactions, as shown above.

We attempted something similar with Llama on IBM Watson AI, but it was not successful. In the current API design for IBM Watson AI foundation models, the entire message must be sent in the 'input' field of the API body, and there is no clear separation of roles (system, user, assistant) within the message.

We have contacted IBM support and were advised to submit an enhancement request, as the option to use historical context is not available in the APIs of the IBM Watson AI foundation models and is not included in the roadmap.

Could you please consider adding this functionality? It is crucial for applications that will be developed using the IBM Watson AI foundation models.

Thank you.