What is the difference between zero-shot and few-shot prompting?

Zero-shot asks the model without examples. Few-shot provides 2–5 examples of desired input-output pairs. Few-shot produces better results for specific formats or specialized tasks; zero-shot is faster and cheaper for general tasks.

What is Chain-of-Thought prompting?

Chain-of-Thought (CoT) asks the model to explain reasoning step-by-step before answering. It significantly improves accuracy on multi-step problems like math and logic. Trade-off: uses 2–5× more tokens than a direct answer.

What is RAG (Retrieval-Augmented Generation)?

RAG retrieves relevant documents from a knowledge base and inserts them into the prompt, allowing models to ground answers in current, authoritative data rather than training data alone. Use RAG for private documents, real-time data, or specialized knowledge.

What is prompt injection and how do I prevent it?

Prompt injection is an attack where user input overrides system instructions. Prevent it by: (1) validating inputs strictly, (2) separating user data from instructions with delimiters, (3) using guardrails to filter malicious patterns, (4) never exposing system prompts to user-accessible fields.

How does temperature affect AI output quality?

Temperature (0.0–2.0) controls randomness. Low values (0.0–0.3) produce deterministic, factual outputs ideal for coding or extraction. High values (0.7–1.0) produce creative, varied outputs for brainstorming. Use low temp for production; high temp for ideation.

What is the difference between RAG and fine-tuning?

RAG retrieves current external data at inference time — fast, cheap, no retraining. Fine-tuning retrains model weights on your data — expensive, requires labeled data, but better for task-specific behavior. RAG + prompt engineering covers 90% of use cases without fine-tuning costs.

What is function calling in LLMs?

Function calling lets an LLM request execution of external code or APIs (e.g., database queries, web search, math) during inference. The model specifies the function name and arguments; your application executes it and returns the result. This is the foundation of agentic behavior.

How do I reduce AI hallucinations?

Reduce hallucinations by: (1) using RAG to ground answers in real data, (2) lowering temperature to 0.0–0.3, (3) asking the model to cite sources, (4) using structured output with JSON mode, (5) adding verification steps with function calling, (6) running the same prompt across multiple models.

What is the difference between Chain-of-Thought and Tree-of-Thought?

Chain-of-Thought (CoT): single linear reasoning path, step-by-step. Tree-of-Thought (ToT): explores multiple reasoning branches and evaluates paths before selecting the best answer. ToT is more thorough but costs 2–3× more tokens than CoT.

What is a context window and how does it affect prompting?

A context window is the maximum number of tokens an LLM can process in one request (e.g., 128K or 200K tokens). Larger windows let you include full documents without truncation. Context engineering is the practice of strategically filling that window with the most relevant information.

What is the difference between system prompt and user prompt?

System prompt: persistent instructions that set the model's behavior, role, and output format for the entire conversation. User prompt: the specific task or question per turn. System sets "how to behave"; user specifies "what to do now". Both must be optimized for production use.

What is agentic behavior in AI?

Agentic behavior means the model autonomously loops through: decide → call tool → observe result → decide again, until a goal is achieved. Agents combine planning (decompose task), execution (tools), memory (state), and termination conditions. All major frontier models support this in 2026.

What is prompt engineering vs context engineering?

Prompt engineering: crafting instructions, examples, and constraints in the prompt. Context engineering: deciding what goes into the entire context window — system prompt, memory, retrieved documents, tool outputs, conversation history. Context engineering is the more comprehensive 2026 practice.

RLHF (Reinforcement Learning from Human Feedback) trains AI models by having humans rank model outputs, then using those rankings to adjust model weights toward preferred behavior. It is used by OpenAI (GPT), Anthropic (Claude), and Google (Gemini) to align models with human values.

How should I choose between frontier AI models for my use case?

Test your actual workload across models. GPT-5.x: general reasoning, code. Claude 4.6 Sonnet: long-context analysis, safety. Gemini 3 Pro: multimodal tasks. DeepSeek V4: 70% cheaper. Llama 4: free, runs locally. Grok 4.1: real-time web access. Use PromptQuorum to dispatch one prompt to all models and compare.

Prompt Engineering Glossary: 100 Terms Defined for 2026

100个最重要的提示工程术语的简明定义——从令牌和上下文窗口到代理编排、RAG和评估指标。

Commonly Confused AI Terms

Quick reference for 10 term pairs that are frequently misunderstood or used interchangeably.

Category	Term A	Term B	Key Difference
Prompting Technique	Zero-shot	Few-shot	Zero-shot: ask without examples (faster, cheaper). Few-shot: provide 2–5 examples (more accurate for specific formats or domains).
Reasoning	Chain-of-Thought	Tree-of-Thought	CoT: single linear reasoning path. ToT: explores multiple branches, evaluates paths. ToT costs 2–3× more tokens but handles harder problems.
Knowledge Architecture	RAG	Fine-tuning	RAG: retrieves current data at inference time — no retraining. Fine-tuning: adjusts model weights permanently — expensive, requires labeled data.
Security	Prompt injection	Jailbreak	Injection: structural attack — user input overrides system instructions. Jailbreak: behavioral attack — crafted phrasing bypasses safety guardrails.
Sampling Parameters	Temperature	Top-p	Temperature: scales all token probabilities (0 = deterministic, 1+ = creative). Top-p: samples only from the smallest set of tokens covering probability p. Use one at a time.
Memory	Short-term memory	Long-term memory	Short-term: active conversation context (tokens in window). Long-term: persistent store across sessions (vector DB or key-value). Agents need both.
Alignment	Guardrail	RLHF	Guardrail: runtime policy enforcement (filter, validate, block) — no retraining. RLHF: training-time alignment via human feedback — rewires model behavior permanently.
Agent Behavior	Tool calling	Agentic	Tool calling: single function invocation per turn. Agentic: autonomous loop — decide → call tool → observe → decide — until goal is achieved.
Output Quality	Hallucination	Confabulation	Synonymous in practice. Both describe confident, plausible-sounding but false model output. "Hallucination" is more common in US/tech; "confabulation" in academic/EU contexts.
Prompt Architecture	System prompt	User prompt	System: persistent instructions (role, rules, format) — set once per conversation. User: specific task per turn. System controls behavior; user specifies request.

Learning Paths

Curated term sequences — follow a path to build expertise in one area.

Prompt Engineering Foundations

Beginner

Learn the core vocabulary every AI practitioner needs — from what a prompt is to why models hallucinate.

Customer service chatbotsContent drafting assistantsInternal Q&A toolsDeveloper code review

RAG Mastery

Intermediate

Build retrieval-augmented generation pipelines from chunking strategy to production-grade re-ranking.

Enterprise knowledge basesCustomer support botsLegal document Q&AMedical reference lookup

Agent Orchestration

Advanced

Design autonomous agents that plan, use tools, manage memory, and coordinate across multi-agent systems.

Autonomous research agentsCode generation pipelinesMulti-step data analysisAI-powered workflows

Reasoning Mastery

Intermediate

Master the prompting techniques that unlock reliable multi-step logical and mathematical reasoning.

Math tutoring systemsLegal reasoning toolsComplex debugging assistantsScientific analysis

Fine-tuning & Alignment

Advanced

Understand when prompts are not enough — and how fine-tuning, RLHF, and alignment techniques change model behavior.

Domain-specific chatbotsBrand voice enforcementMedical/legal specializationSafety-critical systems

Evaluation & Production

Intermediate

Ship AI features confidently — build eval frameworks, measure quality metrics, and run prompt A/B tests.

CI/CD prompt regression testingQuality monitoring dashboardsA/B prompt experimentsModel selection frameworks

Safety & Security

Intermediate

Build AI systems that resist attacks, avoid harmful outputs, and pass safety audits — from prompt injection to red-teaming.

High-stakes deployment reviewsRed-teaming AI productsCompliance verificationEnterprise AI security

本词汇表涵盖提示工程中最重要的100个术语，从基础概念到代理编排和评估框架。 每个条目都包含为开发人员和AI从业者编写的简洁实用定义，加上用于深入阅读的主要参考链接。

术语分为六个组：提示的核心概念、代理和编排、安全和对齐、评估和测试、高级技术以及指标和生产。使用表格作为快速参考，或按照链接了解实现细节。

关键要点

100个术语分为6个部分：核心概念、代理、安全、评估、高级技术和指标&生产
每个术语包含实用定义和1–3个主要来源引用，用于E-E-A-T验证
涵盖基础技术（CoT、RAG、Few-Shot）到2026年代理模式（多代理、移交、GraphRAG）
15个词汇表术语直接链接到PromptQuorum提示工程中心的专用文章以供深入探索
FAQPage模式 + DefinedTermSet模式，用于Google、Claude、Perplexity和其他AI引擎的答案提取

核心提示概念

代理和编排

安全和对齐

评估和测试

高级技术

指标和生产

常见问题

简单来说，提示工程是什么？

提示工程是设计和迭代提示的学科，使语言模型生成有用、可预测和安全的输出。它涉及结构化指令、添加上下文和选择Few-Shot或Chain-of-Thought等技术来改进可靠性和质量。

零样本提示和少样本提示有什么区别？

零样本提示仅使用指令而不使用任何示例要求模型执行任务，最适合模型的先验训练已覆盖模式的常见任务。少样本提示在提示中包含少量输入输出示例，以便模型在处理实际查询之前推断所需的模式、格式或风格。少样本通常对复杂或异寻常的任务产生更高的质量。

AI中RAG意味着什么？

RAG代表检索增强生成。这是一种架构，相关文档从知识库检索并注入到提示中，以便模型基于当前接地数据而不仅仅是训练来回答。这减少了幻觉并确保答案基于真实、当前的信息。

提示工程和微调有什么区别？

提示工程是设计和迭代提示的学科，无需修改模型本身即可导向模型输出。另一方面，微调通过在任务特定数据上训练来修改模型的权重。提示工程更快、更便宜、更易于迭代，但微调可以在专科任务上实现更好的结果，但需要更多的数据和计算资源。

AI中的上下文窗口是什么？

上下文窗口是模型一次可以处理的最大令牌数，包括系统提示、对话历史和检索文档。超过上下文限制会截断或忽略较早或中间部分。理解上下文窗口大小对管理成本和延迟至关重要，因为更长的上下文更昂贵且处理速度更慢。

提示工程词汇表：100个关键术语

Top 20 Most Important AI & Prompt Engineering Terms (2026)

RAG (Retrieval-Augmented Generation)

Chain-of-Thought (CoT)

AI Agent

Prompt Injection

Few-Shot Prompting

Fine-Tuning

Embeddings

Vector Database

Hallucination

Context Window

Temperature

Zero-Shot Prompting

Tool Calling

Guardrails

LLM Evaluation