Terminologies
In the field of prompt engineering, there are many terminologies. This page tries to summarize them add a short and easy explanation about each, so that it becomes easy to remember and stays in memory longer.
-
step-by-step prompting - here the prompt already contains the step by step method to be followed.
-
Chain-of-Thought prompting - here the prompt is extended with additional text such as - 'Think step by step'. This is mostly zero-shot - without no examples provided , or few-shot where multiple example are provided.
-
prompt chaining - Here the task is split into multiple steps and output of first step is sent as a prompt to next step.
-
ReAct - Reason and Act Reasoning - The verb here means, the analysis performed by the LLM to take a certain path in solving an problem. Act - The action the LLM performs based on the reasons.
The action of thinking about something in a logical, sensible way.
This is something what the LLM solely performs.
Reason (taken from the chain of thought) -> act -> observe -> iterate until all steps in the chain are completed.
Reason to Act - Reasoning to finally do some action. and Act to Reason - Performing actions to enhance reasoning. This is where the tool usage comes in.
-
Prompt Sensitivity - Even a small change to the prompt can lead to different answers from LLMs.
-
heuristic - It's a shortcut for thinking. Intuition, gut feel, etc.
-
Agent Strategy - Agent strategy means what LLM to use, what tools to use, which prompting strategy, etc.
Agent is the LLM plus the entire system surrounding it - Chat bot, tools, databases, memory, etc.