How to make AI work for you

Critics of generative AI often discuss the concept of accuracy, when deciding whether or not to use AI.

Though hallucinations, i.e. inaccurate output or more accurately, an incorrect set of outputs, are rendered especially when larger prompts and output text/image/video are displayed, this still doesn't make generative AI a bad tool.

There are ways to fix that and one of them is prompt chaining.

?????????? ?????? ???????????? ???? ??????????

In essence, prompt chaining is using shorter prompts, where one output serves as the next input which serves the next output, etc.

This is very much similar to the Tree-of-Thought (ToT) used in LLM reasoning, itself similar to the decision tree method, though both aren't the exact same thing.

When prompting, we have four choices really:

1. Input-Output Prompting (IOP)

2. Chain-of-Thought Prompting (CoTP)

3. Self-Consistency with CoTP (CoTP-SC)

4. Tree of Thought (ToT)

Input-Output Prompting (IOP)

Clear prompt (input), clear output.

We write something like: "Write me a poem about birds," we receive a poem about birds. Nothing special, no specific reasoning.

Chain-of-Thought Prompting (CoTP)

Input is made of 1-2 steps and the output.

For example: "I have 3 candidates which I want to hire but I only have 2 spots left, how do I choose who is best? Here are their credentials, please explain how I should approach this."

Self-Consistency with CoTP (CoTP-SC)

Input leads to sub-steps (similar to a decision tree) with each leading to another step and 1-3 outputs (for example), of which the majority vote wins.

For example: "I have three expansion plans laid in front of me, each one consisting of X, Y and Z but I can only choose one of these expansion plans in order not to stretch too thin. Can you help me with choosing the best one, considering my current situation and where I want to be in 12 months?"

Tree of Thought (ToT)

The prompt leads to an output which serves as the input for the next, moving forward, continuing the same reasoning and creating a more accurate output as we wanted in the first place, where all steps lead to a single output, using "the garbage" (i.e. the outputs which made little sense) as "fertiliser" for our end output.

Why we should chain our prompts

Though ToT may give a more accurate answer overall, it is not sequential, that is: it may be completely discrete doesn't take into account certain aspects, and does not reason by step-by-step to reach the outcome, that is it is more suitable for planning, summarising, coding and/or debugging than say certain decision-making where you must have a clear-cut choice.

By running a prompt chaining process, we ensure that our auditing would not only be brief but that we could copy-paste the method moving forward, as opposed to receiving different (and sometimes conflicting) outputs, which could become problematic if we have a long task ahead of us.

By chain prompting, we also ensure the system will not need to extrapolate its reasoning method, which consumes computing power which leads to hallucinations and eventually, a pause of a couple of hours in the use of the system and losing our trail of thought.

Are you ready to revolutionise your business processes with AI and unlock unprecedented efficiency and growth?

Contact me to explore how AI can transform your business landscape and discuss tailored AI strategies that align with your unique business goals.

______________________________________________________________________________________

Have you seen Appenue.AI's latest white paper on AI implementation?

Download it here.

#chainprompting #promptengineering #generativeai #genai #ai