3) Prompt Engineering

Vocabulary flashcards based on lecture slides covering key concepts, parameters, techniques, and security aspects of Prompt Engineering.

Prompt Engineering

The process of developing, designing, and optimizing prompts to enhance the output of foundation models (FMs) for specific needs.

Instructions (Prompt Component)

The component of a prompt that defines a specific task for the model to do, including descriptions of how the model should perform.

Context (Prompt Component)

External information included within a prompt to guide the model's response.

Input Data (Prompt Component)

The specific input provided within a prompt for which a response from the model is desired.

Output Indicator (Prompt Component)

An element in a prompt that explicitly specifies the required output type or format.

Negative Prompting

A technique where explicit instructions tell the model what not to include or do in its response, helping to avoid unwanted content, maintain focus, and enhance clarity.

System Prompts

Prompt instructions that define how the foundation model should behave and reply across interactions.

Temperature

A parameter ranging from $0$ to $1$ that controls the creativity of the model's output; low values (e.g., $0.2$) produce conservative, repetitive, and focused responses, while high values (e.g., $1.0$) yield diverse and creative responses.

Top P

A sampling parameter ranging from $0$ to $1$ that restricts candidate words based on cumulative probability; low values (e.g., $0.25$) consider only the top $25\%$ most likely words for coherence, while high values (e.g., $0.99$) consider a broader range for diversity.

Top K

A parameter that limits the pool of probable next words to a specified count $K$; low values (e.g., $10$) create a more coherent response with fewer probable words, while high values (e.g., $500$) produce more diverse and creative output.

Stop Sequences

Specific tokens specified in a prompt configuration that signal the model to immediately cease generating further output.

Prompt Latency

The time it takes for a model to respond, which is affected by model size, model type (e.g., Llama vs. Claude), input token count, and output token count, but is not impacted by Top P, Top K, or Temperature.

Zero-Shot Prompting

Presenting a task to the model without providing any examples or explicit training for that specific task, relying entirely on the model's pre-trained general knowledge.

Few-Shots Prompting

Providing examples of a task within the prompt to guide the model's output; providing only one example is referred to as one-shot or single-shot prompting.

Chain of Thought Prompting

Dividing a complex task into a sequence of explicit reasoning steps (often aided by phrases like "Think step by step") to produce more structured and coherent outputs.

Retrieval-Augmented Generation (RAG)

A technique that combines a model's generative capability with external data sources by augmenting the prompt with relevant information retrieved from external databases.

Prompt Templates

Standardized structures that simplify prompt generation, process user inputs, orchestrate interactions between foundation models, action groups, and knowledge bases, and can be used with Bedrock Agents.

Prompt Template Injections

A security vulnerability or attack (such as an "Ignoring the prompt template" attack) where a user enters malicious inputs designed to hijack prompt instructions and induce unauthorized or harmful responses.