Need for Multi-agent setup

💡 A single Large Language Model (LLM) may not help you completely to arrive at the business solution. Sometimes you have to use one LLM for one task, another LLM for a different task. If one model is good in text generation you can ask this LLM to break your query through a chain of thought prompting. Another LLM can execute these set of actions from previous prompts for different tasks.

➡️ For example, you get a query from consumer. First LLM does the tagging of the query into the respective category, let's say "product packaging issue". Then you use the second LLM to resolve the query and this LLM can use different tools such as company's website links to provide customised solutions.

📈 Sequence of LLMs is something that we call an Agentic framework today. A given task (or a problem) is broken into sub-problems. LLMs (or Agents with language understanding and task execution capability) execute these sub-problems independently. These LLMs work together to solve the entire problem.

🤔 So where does the challenge lie now? Is it in choosing the LLM or in breaking the problem into sub-problems? The challenge lies in each step you take. You need to be creative and thoughtful about what to do and at what time. How do you choose the right LLM? What style of thinking do you want to provide to the LLM through prompts or instructions? What steps would you break your problem into?

📖 At every step/iteration you have to play with the LLMs, use some sequence of steps or the other to solve the problem. It is like you have the chess pieces with you(LLMs) and you know what the pieces do(what the LLMs do) but you learn the sequences that best fits the combination towards winning.

✨ Building Generative AI projects now boils down to choosing LLMs, how you orchestrate them, how you make them specialized in a sub-problem. And in doing so, you move towards your end-goal which is solving the problem in hand.

👇 Share your learning from Generative AI projects and how did you orchestrate LLMs in your case?