This repository provides a streamlined AI terminology guide. It lists key terms and their common alternatives without deep explanations so you can quickly identify prevalent AI concepts. Use it to spot gaps in your knowledge and discover new terms to explore.
💡 See ranking methodology, summary and learning path recommendations at the end of the document.
- Prompt [1/5] |
Natural language instruction, User query- Natural language instruction or question given to an AI model - LLM [2/5] |
Large Language Model, Generative language model, Neural language model- AI model trained on massive text data to understand and generate human-like responses - Output [1/5] |
Response, Generation, Completion- Response generated by an AI model to a given prompt - Token [2/5] |
Subword, Text unit- Smallest unit of text (word, subword, or character) that an LLM processes - Prompt Engineering [2/5] |
Prompt design, Instruction crafting- Art of crafting effective instructions to guide AI models toward desired outputs - Context Engineering [3/5] |
Knowledge management, Context design- Designing and managing information systems provided to LLMs for reliable task completion
- Instruction Following [1/5] |
Task execution, Directive understanding- LLM's ability to understand and execute directions in natural language - Conversation History [1/5] |
Message history, Context log, Dialogue record- Record of previous messages exchanged with an LLM - Turn [1/5] |
Exchange, Interaction cycle- Single exchange in a multi-turn conversation between user and LLM - Response Generation [2/5] |
Text generation, Output creation- Process of LLM creating output based on input
- Zero-Shot Prompting [2/5] |
Direct prompting, No-example prompting- Asking LLM to perform task without providing examples, relying entirely on pre-trained knowledge - One-Shot Prompting [2/5] |
Single example prompting, Single demonstration- Providing exactly one example to clarify the task for the model - Few-Shot Prompting [2/5] |
In-context learning, Multi-shot learning, Demonstration-based prompting- Providing a handful of examples (typically 2-5) to demonstrate desired pattern - In-Context Learning (ICL) [3/5] |
Contextual learning, Learning from examples- LLM's ability to learn from examples provided within prompt - System Prompt [2/5] |
System message, Initial instruction- Initial instructions defining LLM's behavior, role, and constraints - User Message [1/5] |
User input, Query- Instruction or query from end user - Developer Message [2/5] |
Application-level instruction- High-level instructions from application developer (takes priority)
- Clarity [1/5] |
Explicitness, Unambiguity- Making instructions unambiguous and specific - Specificity [1/5] |
Precision, Detail provision- Providing exact details about desired output format, constraints, and requirements - Context Setting [2/5] |
Background provision, Contextual framing- Including relevant background information in prompt - Example Selection [3/5] |
Demonstration curation, Example choice- Choosing diverse, high-quality demonstrations for few-shot learning - Iterative Refinement [2/5] |
Prompt iteration, Incremental improvement, Progressive refinement- Continuously improving prompts based on results - Chain of Thought (CoT) [3/5] |
Step-by-step reasoning, Sequential decomposition, Explicit reasoning- Asking LLM to reason step-by-step through problem by articulating intermediate reasoning steps - Few-Shot Chain of Thought [3/5] |
Demonstration-guided reasoning, CoT with examples- Combining few-shot examples with step-by-step reasoning - Zero-Shot Chain of Thought [3/5] |
Implicit reasoning, CoT without examples- Using prompt cues like `let's think step by step" without examples - Emotion Prompting (EmotionPrompt) [2/5] |
Psychological prompting, Motivational instructions- Adding emotional/psychological stimuli ("This is critical," `Take a deep breath") shown to improve benchmark performance by 8%+
- Prompt Template [3/5] |
Reusable structure, Prompt skeleton- Reusable prompt structure with placeholders for variables - Prompt Caching [4/5] |
Segment caching, Context reuse- Storing frequently used prompt segments for efficiency - Prompt Versioning [3/5] |
Version control, Iteration tracking- Managing different iterations of prompts - Role-Playing [2/5] |
Persona assignment, Character prompting, Role prompting- Instructing LLM to assume specific persona or role for more context-specific, relevant responses - Identity Definition [2/5] |
Behavior specification, Purpose definition- Establishing LLM's purpose, communication style, and goals - Audience Calibration [2/5] |
Audience Persona Pattern, Audience Persona- Tailoring the language and detail of output to audience's knowledge level - Prompt Patterns [3/5] |
Reusable templates, Prompt recipes- Cataloged, reusable prompt structures for specific tasks - Constraint Specification [2/5] |
Boundary definition, Limitation setting- Explicitly defining limits and requirements for outputs - Pseudo-Code Prompting [3/5] |
Code-style instructions, Algorithmic prompting, Structured logic prompting- Prompting using pseudo-code format with control flow syntax, providing structural clues through code comments and docstrings - XML/Markup Prompting [3/5] |
Tag-based prompting, Hierarchical markup prompting- Using XML-style tags to structure prompts and expected outputs, enabling grammar-constrained interaction with fixed-point semantics - JSON Prompting [3/5] |
JSON-style prompts- JSON-style prompts for structured data generation
- Output Format Specification [2/5] |
Format definition, Structure specification- Defining exact structure of desired response - Structured Output [3/5] |
JSON generation, Schema-constrained output- Constraining model to generate JSON or other structured data - JSON Schema [3/5] |
Schema validation, Format enforcement- Specifying exact JSON structure model must follow - Validation [3/5] |
Output checking, Compliance verification- Checking output meets business logic and schema requirements - Prefilling (Output Prefixing) [2/5] |
Response priming, Output seeding, Partial completion- Starting the assistant's response with the desired format or opening tokens to guide the model toward the expected structure (e.g., starting with{for JSON) - Grammar-Constrained Decoding (GCD) [5/5] |
Grammar-guided generation, Formal grammar prompting, Syntax-constrained decoding- Technique that restricts LLM token generation to only valid tokens according to a formal grammar (BNF, CFG), ensuring 100% syntactically correct outputs for domain-specific languages (SQL, JSON, PDDL, SMILES)
- Least-to-Most Prompting [4/5] |
Progressive problem solving, Scaffolded reasoning- Breaking complex problem into simpler subproblems by first breaking down then solving each sequentially, reformulating individual steps to restate previously concatenated results - Decomposed Prompting (DecomP) [4/5] |
Task decomposition, Problem breakdown, Factored decomposition, Modular approach- Breaking tasks into modular components and assigning them to appropriate handlers that can further decompose, solve with simple prompts, or use functions - Plan-and-Solve Prompting [4/5] |
Planning-first reasoning, Two-phase reasoning- Enhancing zero-shot CoT by introducing intermediate planning phase before problem-solving, addressing missing step errors by first planning approach then executing - Self-Consistency [4/5] |
Multiple reasoning paths, Ensemble reasoning, Majority voting- Sampling multiple reasoning paths and selecting most consistent answer through voting, improving reliability over single path reasoning - Analogical Reasoning [3/5] |
Analogy-based prompting, Example-based inference- Using analogies to guide LLM thinking by relating current problem to similar known scenarios - Tree of Thought (ToT) [5/5] |
Branching reasoning, Path exploration- Exploring multiple reasoning paths as tree structure where nodes represent partial solutions, allowing backtracking when paths are unlikely to succeed - Graph of Thought (GoT) [5/5] |
Network reasoning, Interconnected logic- Reasoning across interconnected knowledge as a graph where nodes can reference multiple prior nodes - Abstract Reasoning [4/5] |
High-level thinking, Conceptual reasoning- Working with high-level concepts rather than concrete details - Complexity-Based Prompting [4/5] |
Adaptive depth, Problem-scaled reasoning- Adjusting reasoning depth based on problem complexity, allocating more steps for harder problems - Comparative Reasoning [3/5] |
Relational thinking, Comparison logic- Understanding relationships between concepts through comparison - Skeleton-of-Thought (SoT) [4/5] |
Outline-first generation, Hierarchical decomposition generation- Two-stage approach: first generating skeleton outline, then expanding details in parallel, reducing latency for long-form outputs while maintaining structure - Chain of Density (CoD) [3/5] |
Iterative densification, Entity-dense summarization- Iteratively rewriting summaries to add missing salient entities while maintaining same length, increasing information density without increasing token count - Hint Chain Prompting (HoPC) [4/5] |
Zero-shot pseudo code reasoning, Hint-guided decomposition- Decomposing questions into sub-questions with general pseudo code hints as a chain, executed step-by-step as semantic code interpreter for clearer reasoning `mind map" - Step-Back Prompting [4/5] |
Abstraction-first reasoning, Principle-based reasoning- Evoking reasoning via abstraction by first taking a `step back" to identify high-level concepts and principles before solving the specific problem - Thread of Thought (ThoT) [4/5] |
Sequential context processing, Iterative analysis- Methodically processing, summarizing, and analyzing extended contexts in manageable parts by maintaining a `thread" of key information, resulting in more accurate responses for long documents - Recursion of Thought (RoT) [5/5] |
Recursive decomposition, Divide-and-conquer reasoning- Tackling context length limitations through recursive decomposition using divide-and-conquer approach, particularly effective for large-scale tasks and multi-digit arithmetic - Faithful Chain-of-Thought (CoT) [5/5] |
Verifiable reasoning, Grounded step-by-step- Ensuring LLM reasoning truly reflects the path to the answer, increasing trust and interpretability by guaranteeing final answers derive directly from the reasoning chain
- Self-Critique [4/5] |
Self-evaluation, Output assessment- LLM evaluating its own output for flaws - Self-Refine [4/5] |
Iterative self-improvement, Self-correction, Feedback-driven refinement- Iterative process of LLM improving its own output through multiple rounds - Self-Calibration [4/5] |
Confidence assessment, Uncertainty estimation- LLM assessing confidence in its responses and adjusting certainty levels - Self-Feedback [4/5] |
Internal feedback, Self-assessment loop- Using LLM's own evaluation to drive refinement - Iterative Refinement [3/5] |
Multi-round improvement, Cyclic enhancement- Multiple rounds of improvement with feedback loops - Reversing Chain-of-Thought (RCoT) [5/5] |
Backward reasoning, Hallucination detection, Reverse verification- Reconstructing problem from the answer backwards to detect hallucinations and verify reasoning consistency - Contrastive Chain of Thought (CCoT) [4/5] |
Valid/invalid reasoning, Negative reasoning demonstration- Generating both valid AND invalid reasoning paths (explaining why invalid one is wrong) before final answer, significantly reducing logical fallacies - Chain-of-Verification (CoVe) [4/5] |
Verification questioning, Self-verification- Having model generate verification questions to evaluate its initial response, then answering these questions to refine final output - Self-Verification [3/5] |
Answer checking, Correctness validation- Model verifies its own answers against constraints and requirements - Cumulative Reasoning [4/5] |
Incremental knowledge building, Layered reasoning- Building reasoning incrementally where each step adds to cumulative understanding
- Automatic Prompt Generation [5/5] |
Automated prompt creation, Algorithm-based generation- Using optimization algorithms to generate prompts - Automatic Prompt Optimization (APO) [5/5] |
Self-optimizing prompts- Refining prompts without manual intervention using gradient-based or search methods - MIPRO [5/5] |
Multi-prompt Instruction Proposal Optimizer, Multi-stage optimization, Instruction tuning- Optimizing multi-stage prompts - Meta-Prompting [5/5] |
Prompt generation, Recursive prompting, Self-prompting- Using LLMs to generate and optimize other prompts - AutoPrompt [5/5] |
Automated trigger token discovery, Gradient-guided prompting- Automatically discovering trigger tokens that elicit knowledge from language models using gradient-based search
- Re-reading (RE2) [3/5] |
Question re-examination, Deliberate re-reading- Enhancing reasoning by explicitly asking model to re-read the prompt before answering, ensuring important details aren't missed and reducing comprehension errors - Rephrase and Respond (RaR) [3/5] |
Question rephrasing, Clarity through reformulation- Asking model to rephrase the prompt before answering, reducing ambiguity and improving clarity by letting model reformulate question in clearer terms - SimToM [4/5] |
Simulated Theory of Mind, Perspective reasoning- Enhancing LLMs' ability to understand and predict human thoughts and actions by simulating theory of mind reasoning
- Knowledge Base [2/5] |
Fact repository, Information repository- Structured repository of domain-specific information - Context Window [2/5] |
Token limit, Context length, Attention window- Maximum amount of information/tokens LLM can consider at once - Working Memory [2/5] |
Active context, Immediate memory- Current context available to LLM during generation - Persistent Memory [4/5] |
External memory, Long-term storage, Knowledge storage- External storage (vector database, knowledge base) outside context window
- Information Retrieval [3/5] |
Data extraction, Document retrieval- Extracting relevant data from external sources - Context Retrieval [3/5] |
Information sourcing, Document fetching- Actively sourcing relevant data from external sources - Context Processing [4/5] |
Information optimization, Signal refinement- Optimizing raw information for maximum signal-to-noise - Chunking [3/5] |
Text segmentation, Document splitting- Breaking large documents into smaller, manageable pieces - Summarization [3/5] |
Content condensing, Key point extraction- Condensing information while preserving key meaning - Compression [4/5] |
Context reduction, Token minimization- Reducing context size while maintaining relevant information - Context Management [4/5] |
Context orchestration, Information selection- Orchestrating what information enters LLM's context window - System 2 Attention (S2A) [5/5] |
Context denoising, Deliberate attention prompting- Asking LLM to regenerate input context removing irrelevant information and bias before answering, addressing the `lost in the middle" phenomenon and sycophancy issues
- Episodic Memory [3/5] |
Event memory, Interaction history- Storing specific events or interactions - Semantic Memory [3/5] |
Fact memory, Knowledge storage- Long-term storage of facts and knowledge - Memory Mechanisms [4/5] |
Memory systems, Retrieval mechanisms- Systems for storing and retrieving information - State Persistence [4/5] |
Information continuity, Session memory- Maintaining information across multiple interactions
- Retrieval-Augmented Generation (RAG) [4/5] |
Knowledge-augmented generation, Context-injection generation- Retrieving relevant external documents before generation - Dense Retrieval [4/5] |
Neural embedding search, Semantic retrieval- Using neural embeddings for semantic similarity search - Sparse Retrieval [3/5] |
Keyword search, Lexical retrieval, BM25 search- Traditional keyword-based search methods - Hybrid Retrieval [5/5] |
Combined search, Semantic-keyword fusion- Combining dense and sparse retrieval methods - Vector Search [4/5] |
Similarity search, Embedding-based search- Searching based on numerical vector representations - Semantic Search [4/5] |
Meaning-based search, Context-aware retrieval- Finding information based on meaning rather than exact keywords - Cascading Retrieval [5/5] |
Sequential search, Multi-stage retrieval- Sequential application of multiple retrieval methods
- BM25 [3/5] |
Probabilistic ranking, TF-IDF variant- Probabilistic retrieval model for keyword-based ranking - Dense vs Sparse Tradeoff [4/5] |
Semantic-keyword balance, Precision-recall tradeoff- Balancing semantic understanding with exact matching - Ranking Fusion [4/5] |
Score combination, Multi-signal ranking- Combining multiple ranking signals - Reciprocal Rank Fusion (RRF) [4/5] |
Harmonic mean ranking, Normalized scoring- Technique for merging ranked retrieval results - Learning-to-Rank (LTR) [5/5] |
Gradient boosting ranking- Machine learning approach to optimize ranking - Reranking [4/5] |
Result reordering, Relevance ranking- Re-ordering retrieved results by relevance - Cross-Encoder [5/5] |
Pair-wise scoring, Document-query ranking- Neural model that scores document-query pairs - Late Interaction [5/5] |
Refinement matching, Secondary ranking- Refined matching using multiple signals
- Knowledge Graph (KG) [4/5] |
Entity-relation graph, Semantic network- Structured representation of entities and relationships - Entity Linking [4/5] |
Reference resolution, Entity mapping- Connecting text mentions to knowledge base entities - Graph-Based Retrieval [5/5] |
Graph traversal, Relational search- Using knowledge graphs for information retrieval - Knowledge Graph Embedding [5/5] |
Graph vectorization, KG representation- Vectorizing knowledge graphs for search
- Vector Embedding [3/5] |
Numerical representation, Semantic vector- Numerical representation capturing semantic meaning - Embedding Model [4/5] |
Encoder model, Text-to-vector converter- Neural model converting text to vector representations - Vector Database [4/5] |
Embedding storage, Vector index- Storage system optimized for vector similarity search - Semantic Similarity [3/5] |
Meaning similarity, Contextual closeness- Measuring meaning similarity between texts - Vector Space [3/5] |
Embedding space, Semantic space- Mathematical space where embeddings reside
- Model Context Protocol (MCP) [5/5] |
Context protocol, Standardized interface- Standard protocol for exposing context to LLMs - MCP Servers [5/5] |
Context providers, Protocol servers- Services exposing structured context through MCP - MCP Clients [4/5] |
Protocol consumers, Application clients- Applications initiating MCP requests - Agent Architecture [5/5] |
Multi-component system, Decision framework- Multi-component system with decision-making capabilities
- Transformer Architecture [5/5] |
Self-attention networks, Modern neural design- Neural network design underlying modern LLMs - Attention Mechanism [4/5] |
Focus mechanism, Selective weighting- Focusing on relevant parts of input - Self-Attention [5/5] |
Intra-attention, Internal relationships- Mechanism for relating different positions within a sequence - Cross-Attention [5/5] |
External focus- Attention mechanism between different input sequences
- Softmax Function [4/5] |
Probability converter, Normalization function- Mathematical function converting scores to probabilities - Probability Distribution [4/5] |
Likelihood representation, Statistical distribution- Mathematical representation of likelihood - Logits [3/5] |
Raw scores, Pre-normalization values- Raw prediction scores before probability conversion - Semantic Relationships [3/5] |
Meaning connections, Semantic links- Meaning connections between concepts
- Inference [3/5] |
Model execution, Prediction generation- Running LLM to generate output - Decoding [3/5] |
Output conversion, Text generation- Converting model outputs into human-readable text - Stochastic Behavior [4/5] |
Random behavior, Probabilistic output- Non-deterministic output generation - Probability Sampling [3/5] |
Likelihood-based selection, Random drawing- Selecting options based on likelihood - Speculative Decoding [5/5] |
Draft-verify generation, Assisted decoding, Parallel token verification- Inference optimization using small draft model to propose multiple tokens, verified in parallel by target model, achieving 2-3x — speedup with no quality loss
- Temperature [2/5] |
Randomness control, Creativity parameter- Controls randomness/creativity of output (0=deterministic, high=creative) - Top-K Sampling [3/5] |
K-truncation, Token filtering- Selecting next token from K most probable options - Top-P (Nucleus Sampling) [3/5] |
Probability threshold, Cumulative sampling- Selecting from tokens with cumulative probability up to P - Greedy Decoding [2/5] |
Maximum likelihood, Deterministic selection- Always selecting highest probability token (deterministic) - Random Sampling [2/5] |
Stochastic selection, Uniform sampling- Randomly selecting from entire token distribution
- Max Tokens [1/5] |
Length limit, Output cap- Maximum length of generated response - Frequency Penalty [2/5] |
Repetition penalty, Word reuse reduction- Reducing likelihood of repeating same words - Presence Penalty [2/5] |
Topic diversity, New content encouragement- Encouraging introduction of new topics - Sampling Parameters [3/5] |
Generation settings, Output controls- Controls affecting token selection during generation
- Model Snapshots [2/5] |
Version pinning, Model versioning- Specific versions of models pinned for consistency - Fine-Tuning [4/5] |
Parameter adjustment, Model training- Adjusting model parameters on specific data - Prompt Tuning [4/5] |
Soft prompting, Token learning- Learning soft tokens to prepend to prompts - Prefix Tuning [5/5] |
Prompt prefix learning, Adapter-based tuning- Learning parameters at each layer of model - LoRA [5/5] |
Low-Rank Adaptation, Parameter-efficient fine-tuning- Fine-tuning only low-rank decomposition matrices of weight layers - QLoRA [5/5] |
Quantized Low-Rank Adaptation- Combining quantization with LoRA for memory-efficient fine-tuning - PEFT [5/5] |
Parameter-Efficient Fine-Tuning, Adapter methods- Techniques for fine-tuning with minimal trainable parameters - Quantization [4/5] |
Model compression, Bit reduction- Reducing model precision (Int8, FP16, 4-bit) to decrease size and increase speed - Int8 Quantization [4/5] |
8-bit integer quantization- Converting model weights to 8-bit integers - 4-bit Quantization [5/5] |
Extreme quantization, Ultra-lightweight models- Converting model weights to 4-bit integers for minimal memory footprint - GGUF Format [4/5] |
GPT-Generated Unified Format- Portable format for quantized models enabling local inference - GPTQ Format [5/5] |
Quantization method for GPU inference- Quantization technique optimized for GPU hardware - AWQ Format [5/5] |
Activation-Aware Quantization- Advanced quantization preserving most important weights - Reasoning Models [4/5] |
Deliberative models, Step-by-step thinkers- Models trained for explicit step-by-step reasoning
- Multi-Agent Systems [5/5] |
Agent teams, Collaborative agents- Multiple specialized LLMs working collaboratively - Agent Orchestration [5/5] |
Agent coordination, Task delegation- Coordinating agents to accomplish complex tasks - Intent Translation [4/5] |
Request conversion, Goal interpretation- Converting user requests into actionable instructions - Agentic Reasoning [5/5] |
Agent decision-making, Autonomous planning- LLM deciding which steps to take to solve problem - Plan Decomposition [4/5] |
Task breakdown, Subtask creation- Breaking user request into sub-tasks - Stateful Agents [5/5] |
Persistent agents, Memory-retaining systems- Agents maintaining persistent state across sessions - Multi-Step Planning [5/5] |
Sequential planning, Dependency management- Complex task sequences with dependencies - Cross-Domain Orchestration [5/5] |
Multi-system coordination, Federated agents- Coordinating multiple specialized systems - Autonomous Systems [5/5] |
Self-directed systems, Unsupervised agents- LLM-based systems operating without human intervention - Collaborative AI [5/5] |
Cooperative systems, Inter-AI collaboration- Multiple AI systems working together - ReAct (Reason + Act) [5/5] |
Interleaved reasoning and acting, Thought-Action-Observation- Prompting paradigm alternating Thought (reasoning), Action (tool call), and Observation (result) steps in a loop for agentic tasks - industry-standard pattern for tool-using agents - Multi-Persona Self-Collaboration (Solo Performance Prompting) [5/5] |
Multiple perspectives, Role diversity- LLM collaborates with itself by adopting multiple personas to tackle complex tasks through diverse viewpoints
- Function Calling [4/5] |
API invocation, Tool calling- Instructing LLM to call external functions with arguments - Tool Use [4/5] |
External API usage, Capability extension- LLM invoking external APIs and services - Tool Definition [3/5] |
Function specification, Capability definition- Specifying available functions and their parameters - Agentic Workflows [5/5] |
Agent-driven processes, Autonomous task execution- Multi-step processes where LLM decides which tools to use - Tool Chaining [5/5] |
Function sequencing, Workflow composition- Sequencing multiple tool calls to accomplish complex tasks - Program-Aided Language Models (PAL) [5/5] |
Code-driven reasoning, Programmatic problem-solving- Separating reasoning from computation by expressing reasoning as executable code (Python) that is sent to programmatic runtime, particularly valuable for complex calculations and iterative processes - Chain-of-Code (CoC) [5/5] |
Code-language hybrid reasoning, Semantic-numeric bridge- Framework combining precision of code execution with flexibility of language-based reasoning by generating and executing mix of code and natural language
- Streaming [3/5] |
Token streaming, Real-time generation- Processing response tokens as they're generated - Batch Processing [3/5] |
Bulk requests, Grouped processing- Handling multiple requests in single API call - Chunked Processing [3/5] |
Segmented handling, Partial processing- Breaking responses into manageable pieces
- Token Budgeting [4/5] |
Token allocation, Context distribution- Managing token allocation across context elements - Provenance Tracking [4/5] |
Source logging, Information tracing- Logging which information was provided to LLM - Scratchpad [4/5] |
Work memory, Intermediate storage- External memory for agents to record intermediate work - Thought Chains [4/5] |
Reasoning trace, Explanation chain- Visible reasoning steps leading to conclusions - Tool-Augmented Generation [5/5] |
Tool-enhanced output, Capability extension- Using external tools during generation
- Fill-in-the-Middle (FIM) [4/5] |
Infilling, Middle completion, Context-aware insertion- Technique where model generates missing code/text given both prefix and suffix context, using special sentinel tokens (PSM, SPM, or AST-aware formats) - Horizon-Length Prediction (HLP) [5/5] |
Lookahead planning for FIM, Adaptive infilling- Advanced FIM variant where model predicts normalized length of missing middle content before generation, improving boundary detection by 24%
- Perplexity [3/5] |
Model uncertainty, Prediction entropy- Measure of how uncertain a language model is about predictions (lower is better) - BLEU [3/5] |
Bilingual Evaluation Understudy- Precision-based metric comparing machine translation to reference translations - ROUGE [3/5] |
Recall-Oriented Understudy for Gisting Evaluation- Set of metrics measuring summary quality via n-gram overlap - METEOR [4/5] |
Metric for Evaluation of Translation with Explicit Ordering- Translation metric accounting for synonymy and paraphrases - RAGAS Score [5/5] |
Retrieval-Augmented Generation Assessment, Automated evaluation- Automated RAG pipeline evaluation measuring retrieval and generation quality
- Faithfulness [5/5] |
Hallucination prevention, Grounding score- Measure of whether generated answer is faithful to retrieved context - Answer Relevance [4/5] |
Response pertinence, Query-answer alignment- Measure of how relevant generated answer is to user query - Context Recall [5/5] |
Retrieval completeness, Information recall- Measure of retrieval system's ability to return all relevant documents - Context Precision [5/5] |
Signal-to-noise ratio, Relevant content density- Measure of proportion of retrieved documents that contain relevant information
- Prompt Evaluation [3/5] |
Performance testing, Quality assessment- Testing prompt effectiveness across tasks - Evals [3/5] |
Automated testing, Benchmark tests- Automated tests ensuring consistent behavior - Observability [4/5] |
Monitoring, Logging- Logging and monitoring LLM behavior
- Hallucination [2/5] |
Fabrication, False generation- LLM generating false information confidently - Hallucination Mitigation [4/5] |
Falsification prevention, Grounding techniques- Techniques reducing false outputs - Grounding [4/5] |
Fact anchoring, Source verification- Anchoring LLM outputs in verified sources - Factual Accuracy [3/5] |
Truth verification, Reality alignment- Ensuring outputs match reality
- Safety Alignment [5/5] |
Value alignment, Ethical constraints- Ensuring outputs follow ethical guidelines - Bias Mitigation [4/5] |
Bias reduction, Fairness improvement- Reducing prejudiced outputs
- Prompt Injection [4/5] |
Adversarial prompting, Prompt hacking- Manipulating prompts to influence outputs adversarially - Jailbreaking [4/5] |
Constraint bypass, Safeguard circumvention- Attempts to bypass LLM safety constraints or guidelines - Prompt Leaking [3/5] |
System prompt extraction, Information disclosure- Form of prompt injection where model is asked to output its own prompt, exposing hidden instructions and intellectual property - Data Poisoning [5/5] |
Training data manipulation, Model contamination- Intentionally corrupting training data to degrade model - Model Collapse [5/5] |
Recursive training degradation, Quality decline- Performance degradation from repeated training on model-generated data - Indirect Prompt Injection [5/5] |
Data-based attack, Second-order injection- Injecting prompts through data (documents, web pages) retrieved by system - Refusal Suppression [4/5] |
Safety constraint elimination, Refusal blocking- Sophisticated jailbreak that explicitly constrains model's ability to refuse unsafe requests by prohibiting apologetic language, disclaimers, and refusal words - Obfuscation [3/5] |
Content filter evasion, Text transformation- Technique evading content filters by modifying restricted words through encoding, typos, character substitution, or phonetic preservation - Token Smuggling [4/5] |
Encoded payload delivery, Hidden instruction embedding- Embedding malicious instructions in encoded or obfuscated form to bypass safety filters - Special Case Attack [4/5] |
Edge case exploitation, Boundary condition attack- Exploiting specific edge cases or boundary conditions where safety mechanisms are weaker
- Separate LLM Evaluation [4/5] |
Security layer, Input analysis- Using separate LLM to analyze user inputs for potential risks before main model processes them, adding extra security layer
- LangChain [4/5] |
LLM orchestration framework, Chain composition- Framework for building applications with language models through chaining patterns - LlamaIndex [4/5] |
Data framework, RAG indexing- Framework for indexing and retrieving data for RAG systems - Ollama [3/5] |
Local inference engine, Model runner- Tool for running large language models locally - vLLM [5/5] |
Inference optimization, High-throughput serving- Inference engine optimized for high-throughput LLM serving - OpenAI Assistants API [4/5] |
Stateful threads, Assistant management- API providing persistent threads and assistant state management - DSPy [5/5] |
Declarative programming, Modular prompting- Framework for declarative optimization of language models
- Constitutional AI [5/5] |
Principle-based AI, Value-driven systems- LLMs following explicit constitution of principles - Value Learning [5/5] |
Human value alignment, Ethics learning- Models learning human values from examples - Interpretability [5/5] |
Model transparency, Explainability- Understanding LLM decision-making processes - Mechanistic Interpretability [5/5] |
Internal mechanics, Circuit understanding- Understanding internal computations - Emergent Capabilities [5/5] |
Spontaneous abilities, Scale-dependent features- Abilities appearing unexpectedly at scale
- Adaptive Prompting [5/5] |
Dynamic prompts, Context-sensitive adaptation- Dynamically adjusting prompts based on context - Dynamic Retrieval [5/5] |
On-the-fly retrieval, Conditional fetching- Changing what information to retrieve mid-task - Continuous Learning [5/5] |
Incremental learning, Ongoing improvement- Systems that improve through interaction - Multimodal Context [5/5] |
Mixed media, Multi-sensory input- Integrating text, images, audio in context - Neuromorphic Architectures [5/5] |
Brain-inspired AI, Biologically-motivated systems- Brain-inspired system designs
- Generated Knowledge Prompting [4/5] |
Knowledge generation, Pre-answer knowledge- Asking LLM to generate relevant knowledge before answering the actual question to improve reasoning - Knowledge Augmentation [4/5] |
External knowledge, Information enrichment- Adding external knowledge to improve reasoning - Knowledge Integration [5/5] |
Multi-source synthesis, Knowledge fusion- Combining multiple knowledge sources - Contextual Knowledge [3/5] |
Context-specific facts, Relevant knowledge- Domain-specific information for tasks - Cross-Domain Knowledge [4/5] |
Transfer learning, Multi-domain insight- Applying knowledge across different fields
| Symbol | Meaning | Description |
|---|---|---|
| [5/5] | Expert/Research | Cutting-edge AI techniques requiring deep expertise |
| [4/5] | Advanced | Complex AI concepts for experienced practitioners |
| [3/5] | Intermediate | Core AI techniques requiring solid understanding |
| [2/5] | Beginner | Foundational AI concepts all users should know |
| [1/5] | Essential | Must-know basics for any AI/LLM interaction |
Terms are ranked based on:
- Conceptual Complexity: Mathematical/theoretical depth required
- Implementation Difficulty: Technical skill needed to apply
- Prerequisites: Required foundational AI knowledge
- Research Maturity: Established vs emerging AI techniques
Beginner Journey:
- Foundation Concepts & Basic Interaction
- Basic Prompting Techniques (Zero-shot, One-shot, Few-shot)
- Simple Parameter Controls (Temperature, Max Tokens)
- Understanding Output and Tokens
- Emotion Prompting for performance gains
- Role-Playing for context-specific responses
Intermediate Journey:
- Advanced Prompting (Chain of Thought, Role-Playing)
- Context Engineering Foundations
- Basic RAG Concepts
- Prompt Templates and Patterns
- Pseudo-Code Prompting
- Chain of Density for summarization
- Re-reading and Rephrase-and-Respond techniques
Advanced Journey:
- Complex Reasoning Strategies (Self-Consistency, Tree of Thought)
- RAG System Architecture
- Vector Databases and Embeddings
- Multi-Agent Coordination
- Fine-Tuning and Quantization
- Skeleton-of-Thought for latency
- System 2 Attention for context optimization
- Fill-in-the-Middle for code generation
- Step-Back Prompting and Thread of Thought
- Chain-of-Verification and Self-Criticism
Expert/Research Journey:
- Cutting-Edge Reasoning (Graph of Thought, Meta-Prompting)
- Advanced RAG Architectures (Hybrid Retrieval, Cross-Encoders)
- Model Architecture Theory (Transformers, Self-Attention)
- AI Security Research (Advanced Attacks, Safety Alignment)
- Frontier AI Techniques (Neuromorphic, Constitutional AI)
- Grammar-Constrained Decoding for perfect syntax
- ReAct pattern for autonomous agents
- Speculative Decoding for 2-3x— speedup
- Horizon-Length Prediction for advanced code generation
- Recursion of Thought for large-scale tasks
- PAL and Chain-of-Code for programmatic reasoning
Specialist Path (Prompt Security):
- Start: Prompt Injection Jailbreaking
- Progress: Prompt Leaking Refusal Suppression
- Advanced: Obfuscation Token Smuggling
- Defense: Separate LLM Evaluation Safety Alignment
Specialist Path (Structured Output Engineering):
- Start: JSON Schema Structured Output
- Progress: Prefilling Pseudo-Code Prompting
- Advanced: XML/Markup Prompting Grammar-Constrained Decoding
- Master: Integration with validation and error handling
Specialist Path (RAG/Retrieval Systems):
- Start: Vector Embeddings Vector Databases
- Progress: Dense Retrieval Hybrid Methods
- Advanced: RAG-Specific Metrics Cross-Encoders
- Optimization: System 2 Attention Context Window Management
MIT License This AI terminology guide is released under the MIT License. You are free to use, modify, and distribute this guide for personal and commercial purposes, with attribution.
Contributions are welcome! If you have suggestions for new AI terminology, improvements to definitions, or corrections. Just submit a pull request!