Prompt Engineering Flashcards
(120 cards)
1
Q
What is Access Logging?
A
The recording of all document access events, including who viewed, edited, or downloaded a file, for auditing and security.
2
Q
What are Adapters in neural networks?
A
Lightweight neural network components inserted into a pre-trained model to enable task-specific fine-tuning without modifying the base model weights.
3
Q
Define API Integration.
A
The ability to programmatically interact with a model or service using an Application Programming Interface.
4
Q
What is an Artifact ID?
A
A unique identifier used to reference and retrieve generated or uploaded documents within systems like Grok.
5
Q
Explain Audit Log / Traceability.
A
A system that records all interactions, prompts, and model outputs for accountability, debugging, and compliance.
6
Q
What are Base Weights?
A
The original parameters learned during the pretraining of a model, prior to any fine-tuning or customization.
7
Q
What do BLEU / ROUGE Metrics measure?
A
Evaluation scores used to measure the quality of text generated by comparing it to reference outputs; BLEU is commonly used in translation, ROUGE in summarization.
8
Q
What is Branching Generation?
A
A model generation mode where multiple parallel possibilities are explored rather than following a linear chain of logic.
9
Q
What is the purpose of a Caching Layer in LLM deployment?
A
A system that stores previously computed model responses to speed up response times and reduce compute costs.
10
Q
Define Chained Reasoning.
A
A multi-step logical process where the model connects individual inferences across multiple parts of the prompt.
11
Q
What does Cloud Drive Integration refer to?
A
The capability of an AI service to connect with platforms like Google Drive, Dropbox, or OneDrive for document access and management.
12
Q
What is a Command-Line Interface (CLI)?
A
A text-based method to interact with a model or document service through scripts or shell commands.
13
Q
What is a Context Window?
A
The maximum number of tokens (words and punctuation) that a model can consider at once while generating a response.
14
Q
What is a Contextual Memory Agent?
A
An LLM-based system designed to retain, recall, and use session-specific information across multiple interactions.
15
Q
What is Cross-document Linking?
A
The ability to reference, connect, or embed information between multiple documents in a system.
16
Q
What is a Custom Metadata Schema?
A
User-defined structures used to attach properties like author, topic, or status to documents for classification or querying.
17
Q
What is Data Augmentation?
A
Techniques for expanding training datasets using modified, paraphrased, or synthesized examples to improve model robustness.
18
Q
Define Dialogue Bias.
A
A setting or characteristic that influences the model’s tone, persona, or role adherence during conversation.
19
Q
What is Document Chunking?
A
The process of breaking a large document into smaller, manageable segments for input into an LLM with context window constraints.
20
Q
What is Document Indexing?
A
The process of organizing and storing documents in a way that allows for semantic search and retrieval.
21
Q
What does Domain Adaptation involve?
A
Adjusting a pre-trained model to perform better on a specific domain or type of data by fine-tuning or modifying prompts.
22
Q
What are Embeddings?
A
Vector representations of words, sentences, or documents that capture semantic meaning for tasks like similarity search or clustering.
23
Q
What is Emotion Bias?
A
A model tuning configuration that encourages responses to align with a specific emotional tone (e.g., optimistic, neutral).
24
Q
What is an Evaluation Benchmark?
A
A standardized dataset and task used to compare the performance of different LLMs or configurations.
25
What does Fallback Behavior refer to?
The model's default response when encountering input outside its training distribution or intended domain.
26
What is the F1 Score?
A statistical measure of a model’s accuracy that balances precision and recall, commonly used in classification tasks.
27
Define Factual Calibration.
Adjusting model outputs toward better factual accuracy using postprocessing or model-level weighting.
28
What is File Path Abstraction?
A logical representation of file location, often used in LLM systems to simulate file directories (e.g., /mnt/data/file.txt).
29
What is a Fine-tuning Objective?
The specific goal or use case for which a model is being fine-tuned, such as classification, summarization, or translation.
30
Define Flat Namespace.
A storage scheme where all files exist at a single level, without folders or hierarchy.
31
What does Frequency Penalty do?
A model parameter that reduces the probability of words being repeated in a single generation.
32
What is Function Calling / Tool Use?
The ability of a model to interact with external tools, APIs, or code execution environments to augment its responses.
33
What is Ground Truth Comparison?
The process of evaluating model outputs against known correct answers to measure performance or accuracy.
34
What is a Hierarchical File System?
A file storage structure where documents are organized in nested directories or folders.
35
What is HTML Output?
A format for displaying structured content using HyperText Markup Language; useful for formatting complex output like tables.
36
Define Human-in-the-Loop (HITL).
A workflow in which humans review or correct model outputs, often used in sensitive or critical applications.
37
What is Incremental Fine-tuning?
A refinement process where new data is continually added to a model's training to improve its performance over time.
38
What does Input Document Parsing involve?
The method by which input documents are analyzed or broken down into parts such as sections, headers, or tables.
39
What is an Instruction Prefix?
A predefined directive prepended to the prompt to shape or control model behavior (e.g., 'You are a helpful assistant...').
40
Define Instruction-tuning.
A fine-tuning method where the model is trained on pairs of instructions and desired outputs to improve alignment with user intent.
41
What is Latency in model responses?
The delay between submitting a prompt and receiving a model response; a critical performance metric in real-time applications.
42
What is a Latent Representation?
A compressed, high-dimensional internal state that captures the semantic features of input data.
43
What does Linearity refer to in model generation?
A generation pattern where the model adheres strictly to a single line of reasoning or narrative progression.
44
What is LoRA (Low-Rank Adaptation)?
A parameter-efficient fine-tuning technique that inserts low-rank matrices into existing layers of a model.
45
Define Markdown Output.
A plain-text markup format that supports structured formatting, often used for documents generated by LLMs.
46
What does Max Tokens refer to?
A generation constraint defining the maximum number of tokens (input plus output) in a single response.
47
What is Memory Persistence?
The ability of an LLM system to retain session-specific or user-specific context between interactions.
48
What is Meta-prompting?
The practice of prompting a model to generate or modify other prompts, typically for optimization or automation.
49
What is Modular Prompt Orchestration?
The use of composable prompt segments or chains to control model reasoning and output structure.
50
What does Model Bias Calibration involve?
Adjusting a model’s responses to reduce or rebalance systemic biases introduced during pretraining.
51
Define Multimodal Bias.
A tuning mechanism to guide output preference toward certain input types such as images or text.
52
What does Multimodal Input/Output refer to?
Support for interacting with multiple types of media (e.g., text, images, audio) as input and/or output.
53
What is Multi-agent Collaboration?
The coordination of multiple LLM instances (agents), each with a specialized role, to complete a task collaboratively.
54
What is Nucleus Sampling (Top-p)?
A token selection strategy that limits generation to a subset of tokens whose cumulative probability meets a threshold.
55
What are Optional Parameters?
User-controlled or experimental settings (e.g., Emotion Bias, Logical Coherence) that adjust generation style or content bias.
56
What does Output Determinism mean?
The consistency of model output when given the same input and configuration; influenced by seed and temperature.
57
What is Overfitting?
A condition where a model performs well on its training data but poorly on new, unseen inputs due to excessive memorization.
58
Define Parameter-efficient Fine-tuning.
Techniques like LoRA or adapters that fine-tune only a subset of model parameters for faster adaptation.
59
What is a Presence Penalty?
A parameter that penalizes tokens that have already appeared in the prompt, encouraging diversity in output.
60
What does Prompt Chaining involve?
The method of linking multiple prompts together so that the output of one becomes the input of the next.
61
Define Prompt Engineering.
The practice of designing and refining prompts to elicit better or more accurate responses from a model.
62
What is a Prompt Formatting Schema?
A structured format expected by the model to guide response generation (e.g., Input: ___ | Instruction: ___).
63
What does Prompt Orchestration entail?
The coordination and management of prompt flows, structures, and dependencies in a multi-step or multi-agent system.
64
What is a Prompt Template?
A reusable prompt framework with placeholders for dynamic content, used to standardize interactions.
65
What is a Random Seed?
A value used to initialize the random number generator in language models, making output reproducible.
66
What is a Repetition Penalty?
A tuning parameter that reduces the chance of repeated phrases or tokens in generated text.
67
What is Retrieval-Augmented Generation (RAG)?
A technique where the model retrieves relevant documents before generating responses to ground its output.
68
What does Role-weighted Prompting refer to?
Assigning different roles to sections of a prompt or to multiple agents, influencing tone and responsibility.
69
Define ROUGE Score.
A metric that measures the overlap between generated and reference summaries, often used in NLP evaluation.
70
What is Sandboxing?
Isolating an LLM or its runtime to prevent interaction with external systems or to safely test prompts and responses.
71
What is Semantic Search?
A method of retrieving documents or data based on meaning and context, often using embeddings.
72
What is a Session Artifact?
A document or output object tied to a specific user interaction session within an LLM platform.
73
What is Session Contextualization?
The use of prior turns or session memory to ground the current response in historical interaction.
74
What does Session Linking refer to?
The ability to reference or build upon interactions from earlier sessions in new conversations.
75
What are Stop Sequences?
Designated tokens or characters used to signal where the model should stop generating text.
76
What is Structured Data Handling?
The model’s capability to accurately process and generate data formats like JSON, XML, or tabular formats.
77
What is a Summarization Objective?
A fine-tuning goal where the model is trained to reduce content while preserving key information.
78
Define Symbolic Reasoning.
Logic-based reasoning involving symbols and rules, as opposed to statistical associations.
79
What is a System Prompt?
An instruction injected at the beginning of a session or prompt sequence to define the assistant’s behavior.
80
What is a Tagging System?
A metadata scheme used to classify, label, or search documents based on their properties.
81
What does Task-Specific Agent Design involve?
Building specialized LLM agents optimized for narrow tasks like bug triage, legal review, or translation.
82
What is Temperature Sampling?
A method for controlling randomness in language generation; lower values yield more deterministic outputs.
83
What are Throughput Constraints?
The limits on how many requests or tokens can be processed per second or per batch.
84
What is Token Biasing?
A generation control setting that promotes or suppresses specific tokens during output.
85
Define Token Billing Granularity.
The level of detail at which token usage is measured for cost calculation.
86
What is a Token Limit?
The maximum number of tokens the model can process in a single interaction (input + output).
87
What does Tool Use / Plugin Execution refer to?
The capacity of the model to invoke external services (like code execution, database queries, or calculators).
88
What is Top-k Sampling?
A sampling method where the model chooses the next token from the k most probable candidates.
89
What is Transfer Learning?
Leveraging knowledge from a model trained on one task/domain and applying it to a related one through fine-tuning.
90
What is a Vector Database?
A storage system that indexes embeddings (vector representations) for fast similarity search and retrieval.
91
What is Version Control?
A feature that allows users to track changes, retrieve prior versions, or revert documents to earlier states.
92
93
What is Format Anchoring?
A prompt engineering technique where the desired output format is specified within the prompt to ensure consistent and structured responses.
## Footnote
Examples include specifying formats like JSON, Markdown, or tables in the prompt.
94
Define Front-Matter Prompting.
A prompting strategy where metadata is embedded at the beginning of a document to influence model behavior or classification.
## Footnote
Common metadata formats include YAML.
95
What is Few-shot Prompting?
Providing the model with a small number of input-output examples in the prompt to demonstrate the task.
## Footnote
This helps guide output formatting or logic.
96
Explain Low-Temperature Precision Prompting.
Using low randomness settings to generate consistent and accurate model outputs, especially useful for technical content.
## Footnote
Temperature settings typically range from 0.2 to 0.4.
97
What does Pre-injection of Examples entail?
Inclusion of one or more complete examples before posing a new input to help ground the model’s response.
## Footnote
This strategy aids in establishing learned patterns.
98
Define Prompt Templates.
Predefined prompt structures with variable placeholders that can be reused across contexts to standardize task instructions.
## Footnote
They improve consistency in responses.
99
What is Prompt Chaining?
A technique where the output of one prompt becomes the input of the next, enabling multi-step reasoning workflows.
## Footnote
This allows for more complex interactions.
100
What does Instruction Prefixing involve?
The practice of explicitly starting prompts with an instructional clause to control tone and response structure.
## Footnote
An example might be, 'Explain this as a lawyer...'.
101
Define System Prompts.
Persistent hidden prompts used to define the model’s role, tone, or behavior across an entire session.
## Footnote
These prompts influence the consistency of responses.
102
What is Meta-Prompting?
Prompting the model to write or optimize other prompts, used for adaptive or agent-based LLM design workflows.
## Footnote
This approach improves the efficiency of subsequent prompts.
103
Define Role-weighted Prompting.
Assigning roles or personas to parts of a prompt or to simulated agents in a multi-agent system.
## Footnote
Roles can include Critic, Engineer, or Advocate.
104
What does Embedding-Aware Retrieval Prompting involve?
Supplying retrieved context within the prompt to ground the model’s answer.
## Footnote
This can be achieved via semantic search.
105
What is Chunk-aware Prompting?
Indicating document segmentation explicitly to improve model understanding of document structure.
## Footnote
An example would be labeling sections as 'Part 2 of 4'.
106
Define Self-reflective Prompting.
Asking the model to assess or critique its own response before finalizing it.
## Footnote
This process improves reliability or completeness of answers.
107
What is Stop Sequence Control?
The use of explicit stop sequences in prompts to limit the model’s output, especially in structured formats.
## Footnote
This is particularly useful for formats like code or JSON.
108
Define Temperature-Aware Prompting.
Designing prompts with an understanding of how randomness affects creativity, verbosity, and stability of output.
## Footnote
Different temperature settings can yield varying levels of output quality.
109
What is Token-efficient Prompting?
Writing prompts that deliver high performance with minimal token usage.
## Footnote
This is achieved by compressing instructions and avoiding filler language.
110
Define Zero-shot Prompting.
Asking the model to perform a task without providing any examples, relying on natural language instruction.
## Footnote
It tests the model's ability to generalize from instructions alone.
111
What is Ambiguity in Prompting?
A condition where prompts contain overlapping or unclear instructions, leading to inconsistent or inaccurate model responses.
## Footnote
This can affect the quality of the output generated by a model.
112
Define Context Overflow.
A situation where the input exceeds the model’s context window, causing earlier parts of the prompt to be truncated or ignored.
## Footnote
This can lead to loss of important information in the response.
113
What is Instruction Overload?
The practice of including too many directives in a single prompt, often reducing clarity and model performance.
## Footnote
Simplifying instructions can enhance the quality of the output.
114
What does a Prompt Style Guide entail?
A standardized set of formatting rules and structures used to ensure consistency across prompts within an organization or project.
## Footnote
This can help maintain uniformity in how prompts are crafted.
115
Explain Prompt Token Budgeting.
The strategy of managing the number of tokens used in prompts to optimize for performance and cost efficiency.
## Footnote
Efficient token usage can lead to better model performance.
116
What is Relevance Filtering?
The process of selectively including only contextually important content (e.g., in RAG) to reduce noise and improve model focus.
## Footnote
This helps in enhancing the relevance of the output.
117
Define Response Drift.
A tendency for models to produce output that gradually strays from the intended task, often due to unclear structure or poor anchoring.
## Footnote
Clear structure can help mitigate this issue.
118
What is Runaway Output?
When a model generates excessively long or irrelevant text due to a lack of defined stop sequences or constraints.
## Footnote
Implementing stop sequences can help control output length.
119
What does Structured Prompting involve?
Organizing prompts using labeled or segmented parts (e.g., Input | Task | Output Format) to improve response reliability.
## Footnote
This structure can help models understand the task better.
120
What is Under-specified Prompting?
Prompts that are too vague or general, causing the model to guess at user intent rather than follow clear guidance.
## Footnote
Providing specific details can enhance the accuracy of the output.
