Google continues to publish and open-source countless generative AI resources to the wider community, whether you’re a business practitioner, software engineer, or AI/ML data scientist. Those looking to develop a strong understanding of advanced prompt engineering techniques and tuning will be well suited to experiment with a free Google Jupyter notebook training on two powerful large language model (LLM) prompting strategies: Chain of Thought and ReAct (Reasoning and Acting) using LangChain. Entitled, “Advanced Prompt Engineering”, the notebook is free to download and use and helps users understand how to use Generative AI technologies on Google Vertex AI.
The Jupyter notebook uses Python and is available for anyone to download and experiment with on GitHub. Other content available within the repo not covered in this post include developer productivity with GenAI, LangChain observability, Vertex AI evaluation services, and Vertex AI foundation tuning.
LLM Prompting Strategies: Chain of Thought and ReAct
The Google notebook teaches two powerful prompting techniques: Chain of Thought and ReAct (Reasoning + Acting). ReAct (and its variants) are the current state-of-the-art prompting technique to improve LLM reasoning while minimizing hallucinations. Chain of thought is a relatively low-effort technique to improve prompt performance and robustness by adding verbal reasoning. The notebook also covers LLM tools/actions, self-consistency, a zero-shot chain of thought, and some basics of how LangChain does ReAct.
Chain of Thought reasoning encourages large language models to get the right answer when they first output text that explains the reason for their answer. Essentially, it boosts reasoning by thinking of intermediary steps to solve a problem.
The ReAct framework leverages Chain of Thought reasoning together with granting access to the LLM with external tools to interact with the “public world” in the form of agents. This is accomplished using LangChain for execution of tasks with tools. Example use cases are for Retrieval-augmented Generation (RAG), interacting with APIs, or chatbots.
This is an example of task-specific tuning to make LLMs more reliable, and useful.
The four parts of the notebook include training on the following:
- Chain of Thought Prompting: Using language descriptions of reasoning to improve LLM outputs.
- Actions, Retrieval, and Tool Use: How LLMs interact with external systems.
- ReAct (Reasoning + Acting) Prompting: Combining the written reasoning descriptions of Chain of Thought prompting with external system interactions.
- Langchain and ReAct: What to expect when using Langchain ReAct agents.
The notebook can be run on Google Colaboratory, Vertex AI Workbench, Visual Studio Code, JupyterLab, or your other favorite Jupyter-compatible notebook environment.
The training is assuming that users have familiarity with large language models, what an LLM is and how they work, basic experience with LLM prompting, and the difference between zero-shot, one-shot, and few-shot prompting.
If you’d like a refresher on the fundamentals of prompt design and prompt tuning, check out Google’s Introduction to Prompt Design documentation.
Additional free advanced training content is available for Google generative AI solutions and concepts such as large language models, image generation, image captioning models, and encoder-decoder architecture.
