Module 3 - Prompt Engineering Techniques

Prompts with examples: Zero-Shot, One-Shot, Few-Shot techniques

  • In-context learning: the LLM learns the task on the spot from the examples you provide
  • Language Models are Few-Shot Learners paper demonstrated that larger models benefit disproportionately more from in-context examples
  • Zero-Shot Prompting - no examples given
  • One-Shot Prompting - 1 example given
  • Few-Shot Prompting - 2 or more examples given

    • Thought-Based Prompting Techniques

  • Chain-of-Thought (CoT) Prompting - guide the AI to break down a problem into sequential, logical steps
  • Automatic Chain-of-Thought - simply tell the AI to go through the answer step-by-step
  • Tree-of-Thought (ToT) tell the AI to consider multiple paths at each step like a decision tree; You can backtrack and explore different paths
  • Thread-of-Thought - stay on one path without branching
  • Iteration-of-Thought (IoT) - after each step, the AI decides whether to continue on the current path, adjust its reasoning, or restart with a new approach.
  • Scratchpad Prompting - tell the AI to write down intermediate steps while working through a problem
  • Step Back and Reflection Prompting - Tell the AI to pause and take a broader view of its current progress before moving forward.
  • Self-Consistency - Tell the AI to generate multiple independent responses. Choose the most common answer.
  • Rephrase and Respond (RaR) - ask the AI to rephrase or restate the input question before generating a response.
  • Echo Prompt Technique - simply adding "Repeat the question before answering it." to your prompt can make the model answer questions more effectively.

    • Automatic Prompt Engineering (APE)

  • Use algorithms or models to automatically optimize prompts.
  • Can save time.
  • Can make prompts that use fewer tokens
  • Use one AI to generate prompt variations; score those candidates using another AI; select the best-performing version.
  • Generate prompts by searching for or by using AI to make 'em

    • Security

  • Prompt injection: Adversarial inputs embedded in user data that attempt to manipulate model behavior.
  • Reliance on untrusted inputs: Automated pipelines that incorporate external data sources need to minimize reliance on untrusted inputs.
  • Reproducibility: Automated prompt optimization may produce prompts that perform well on benchmarks but behave unexpectedly in production contexts.

    • Tool Use

  • Modern LLMs can access live data, perform computations, execute code, and call APIs.
  • Tool calling (also called function calling) prompts the AI to call APIs
  • Zero-Shot Tool Usage via Documentation - simply providing clear tool documentation allows LLMs to use APIs effectively
  • Overuse or misuse: Invoking tools unnecessarily can slow down LLM responses or lead to unexpected results.
  • Security threats: Tool-based prompting increases the risk surface, opening the door for malicious prompt injection or adversarial misuse.
  • Reliability: LLMs may output malformed tool calls or misuse documentation, emphasizing the need for robust prompt design and error handling

    • Retrieval-Augmented Generation (RAG)

  • Retrieval-Augmented Generation (RAG) is a hybrid approach that integrates retrieval mechanisms with generative models.
  • primary goal of RAG is to reduce hallucinations.

    • How RAG Works: A Two-Step Architecture

  • Retrieval: Retrieval component searches a predefined corpus to find relevant information.
  • Generation: The relevant information is included in the prompt so the model can synthesize the information to create a response.

    • Applications of RAG

  • Question Answering: RAG models provide precise answers by retrieving relevant documents and generating responses based on that information.
  • Summarization: By accessing multiple sources of information, RAG creates comprehensive summaries that capture the essence of content.
  • Conversation: RAG enhances dialogue systems by allowing them to pull in real-time information, making interactions more informative and engaging.
  • Content Creation: Writers and content creators leverage RAG to generate articles or reports that are well-informed and relevant to current events or specific topics.
  • Enterprise Knowledge Bases: Organizations use RAG to allow employees to query internal documents, policies, and knowledge bases conversationally.

    • How Does RAG Find Relevant Documents?

    Dense Retrieval

    Keyword-Based Search