Assistant (to the) Regional Planner
Putting the lessons from DeepLearning.AI to use by building an assistant to help with a series of tasks. The assistant converses with the user to get the confirmation before moving through the steps in sequence. The assistant can determine the relevant function for each step of the process and answer questions on the business data. Learning by doing!
My note to keep the assistant on point works very well. It takes three prompts :-)
A few thoughts at the end of the process.
- Testing is no mean feat. Responses from LLMs are not guaranteed to be consistently correct. QA teams will have to make provision for failure and report that the step was tested successfully in all but 2% of the cases :-)
- Unambiguous naming for the tables and fields is going to be important. Applications might need a forced autocorrect ensuring that users using different terms (product, an SKU or an FG ) will not throw the LLM off.
- The words in the prompt matter, which brings up questions how one could validate the quality of prompts to handle real world cases.
Technical Stuff is here https://sharathbobba.wordpress.com/2024/05/16/conversational-assistant/