Concept Fables
把 AI / Agent 概念讲成能记住的故事
这里收录 682 个概念。每个词条先用寓言建立理解过程，故事结束后再揭示定义和隐喻映射。
领域分类
全部概念
682 个
AI agent platform
精修版Agent 运行时
Agent Runtime
Agent 运行时是负责执行智能体循环、管理状态、调用工具、处理权限和恢复错误的运行环境。
读寓言 →
AI agent platform
精修版工具调用
Tool Calling, Function Calling
工具调用是模型以结构化方式请求外部函数、API、数据库或系统能力来完成任务的机制。
读寓言 →
AI agent platform
精修版工具注册表
Tool Registry
工具注册表集中描述可用工具的名称、参数、权限、调用方式和风险等级，供 agent 运行时选择和管控。
读寓言 →
AI application observability and evaluation tool
Weights & Biases Weave
Weights & Biases Weave
Weave is a W&B tool for tracing, evaluating, debugging, and iterating on LLM applications and model-powered workflows.
读寓言 →
AI engineering / data pipeline
合成数据
Synthetic Data
合成数据是由模型、规则或仿真系统生成的数据，用于补充训练、评测、微调或测试场景覆盖。
读寓言 →
AI engineering / fine-tuning
低秩适配
LoRA, Low-Rank Adaptation
LoRA 是一种参数高效微调方法，通过训练小规模低秩矩阵来适配大模型，而不是更新全部模型参数。
读寓言 →
AI engineering / fine-tuning
量化低秩适配
QLoRA, Quantized LoRA
QLoRA 在量化后的基础模型上训练 LoRA 适配器，从而显著降低大模型微调所需显存。
读寓言 →
AI engineering / infrastructure
推理引擎
Inference Engine
推理引擎是负责高效执行模型前向计算、显存管理、批处理、调度和加速优化的软件组件。
读寓言 →
AI engineering / model optimization
蒸馏
Distillation
蒸馏是用大模型或强模型的输出训练较小模型，使小模型以更低成本复现部分能力。
读寓言 →
AI engineering / operations
批量推理
Batch Inference
批量推理是将大量输入集中提交给模型离线处理，适合报表、标注、推荐召回和非实时任务。
读寓言 →
AI engineering / operations
在线推理
Online Inference
在线推理是在用户请求发生时实时调用模型生成结果，重点约束通常是延迟、可用性和并发能力。
读寓言 →
AI platform / cost optimization
精修版模型路由
Model Routing
模型路由根据任务、成本、延迟、质量或权限策略选择调用不同模型或供应商。
读寓言 →
AI platform / cost optimization
精修版语义缓存
Semantic Cache
语义缓存通过判断新请求与历史请求的语义相似度来复用已有回答或中间结果，从而降低成本和延迟。
读寓言 →
AI platform / deployment
推理端点
Inference Endpoint
推理端点是模型服务暴露给应用调用的网络入口，通常包含鉴权、限流、路由和版本管理能力。
读寓言 →
AI platform / infrastructure
模型网关
Model Gateway, AI Gateway
模型网关统一接入多个模型供应商或自部署模型，并提供鉴权、路由、限流、日志、缓存和成本控制。
读寓言 →
AI platform / integration
精修版模型上下文协议
Model Context Protocol, MCP
MCP 是一种让 AI 应用以标准方式连接外部工具、数据源和上下文服务的协议。
读寓言 →
AI product metrics / operations
AI 工作替代率
AI Task Automation Rate
AI 工作替代率衡量原本由人完成的任务中有多少比例被 AI 自动完成或显著减少人工介入。
读寓言 →
AI product metrics / support operations
自动化偏转率
Deflection Rate
自动化偏转率衡量 AI 或自助系统成功处理并避免转人工处理的请求比例，常用于客服、IT 和运营场景。
读寓言 →
AI product pattern
自动驾驶
Autopilot
Autopilot 是让 AI 在明确边界和监控下自主完成任务的产品模式，人工主要负责设定目标、监督和处理异常。
读寓言 →
AI product pattern
副驾驶
Copilot
Copilot 是辅助人完成工作的 AI 产品模式，通常由人做最终判断、编辑或批准。
读寓言 →
AI safety
水印
Watermarking
Embedding detectable signals into AI-generated content to indicate origin.
读寓言 →
AI-native business model
服务即软件
Services-as-Software
Services-as-Software 是用 AI 和自动化把原本由人工交付的服务流程产品化，使客户购买的是业务结果而不只是工具访问权。
读寓言 →
Agent Operating Model
Agent 运营模型
Agent Operating Model
The organizational design for how AI agents are owned, assigned work, supervised, measured, improved, escalated, and governed in production.
读寓言 →
Agent Operating Model
Agent 所有人
Agent Owner
The accountable business or technical owner responsible for an agent's purpose, permissions, performance, risks, and lifecycle decisions.
读寓言 →
Agent Operating Model
职权边界
Authority Boundary
The explicit limit of decisions, actions, spending, data access, and external commitments an AI agent is allowed to make.
读寓言 →
Agent Operating Model
自主级别
Autonomy Level
A classification of how independently an AI agent can recommend, decide, execute, monitor, or escalate work.
读寓言 →
Agent Operating Model
运营指标
Operating Metrics
Quantitative measures used to manage AI agents in production, such as completion rate, escalation rate, error rate, latency, cost, and rework.
读寓言 →
Agent architecture
智能体循环
Agent Loop
The agent loop is the repeated cycle in which an AI agent observes state, reasons or plans, selects an action, executes it through tools or code, and incorporates the result into the next step.
读寓言 →
Agent architecture
执行器
Executor
An executor carries out planned actions by calling tools, running code, retrieving information, or invoking external services, then returns results for evaluation or replanning.
读寓言 →
Agent architecture
规划器
Planner
A planner decomposes a user goal into ordered steps, subgoals, constraints, or task graphs that can be executed and revised as new information arrives.
读寓言 →
Agent capability
动作
Action
A discrete operation an agent can execute in the environment, often through a tool or API.
读寓言 →
Agent capability
计算机使用
Computer Use
模型通过屏幕、鼠标、键盘或 UI 自动化直接操作计算机环境。
读寓言 →
Agent capability
GUI 智能体
GUI Agent
通过视觉识别和界面操作完成软件任务的 Agent。
读寓言 →
Agent capability
网页智能体
Web Agent
在网页环境中搜索、导航、表单填写和执行操作的智能体。
读寓言 →
Agent decision-making
工具选择
Tool Selection
Tool selection is the process by which an agent decides which available tool, function, or API to invoke for a given subtask and with what parameters.
读寓言 →
Agent execution object
运行
Run
A run represents one execution of an assistant over a thread, including model reasoning, tool calls, required actions, and final output.
读寓言 →
Agent execution trace
运行步骤
Run Step
A run step records an intermediate action during a run, such as a model message creation or tool call, enabling inspection of execution progress.
读寓言 →
Agent interface
实时智能体
Real-Time Agent
在低延迟流式交互中理解、规划、调用工具并回应的 Agent。
读寓言 →
Agent interface
语音智能体
Voice Agent
通过语音输入输出与用户实时交互并完成任务的 Agent。
读寓言 →
Agent memory
情景记忆
Episodic Memory
Episodic memory stores specific events, interactions, observations, and temporal experiences, usually with metadata such as time, source, actors, and outcome.
读寓言 →
Agent memory
记忆压缩
Memory Compression
Memory compression condenses long histories, observations, or retrieved materials into shorter summaries or structured facts that fit future context budgets while preserving useful signal.
读寓言 →
Agent memory
语义记忆
Semantic Memory
Semantic memory stores generalized facts, concepts, preferences, and learned knowledge abstracted from individual events rather than preserving the full episode.
读寓言 →
Agent operations / governance
审批流
Approval Flow, Human Approval
审批流是在高风险、外部写入、付费或破坏性动作前要求人工确认的控制机制。
读寓言 →
Agent orchestration pattern
状态图
State Graph
A state graph models an agent workflow as nodes that read and update shared state while edges determine the next execution step.
读寓言 →
Agent protocol
Agent-to-Agent 协议
Agent2Agent, A2A
A protocol direction for interoperability and communication between autonomous agents across systems.
读寓言 →
Agent state
长期记忆
Long-Term Memory
Long-term memory is persistent information saved outside the immediate context window and retrieved later to support continuity, personalization, or accumulated domain knowledge.
读寓言 →
Agent state
工作记忆
Working Memory
Working memory is the short-lived task state held in the current context, including goals, constraints, observations, intermediate artifacts, and recent tool results.
读寓言 →
Agent state / reasoning workspace
草稿区
Scratchpad
A scratchpad is a temporary workspace where an agent records intermediate observations, calculations, plans, or tool results for use during the current task.
读寓言 →
Agent workflow framework
LangGraph
LangGraph
LangGraph is a graph-based framework for building stateful, controllable, multi-step agent workflows with explicit nodes, edges, state, and persistence.
读寓言 →
Alignment
对齐
Alignment
The process of making model behavior better match human intentions, preferences, policies, and safety constraints.
读寓言 →
Alignment
宪法式 AI
Constitutional AI
An alignment approach that uses written principles to critique and revise model outputs, reducing reliance on direct human labels for every example.
读寓言 →
Alignment
直接偏好优化
Direct Preference Optimization, DPO
A preference tuning method that optimizes a language model directly from chosen-versus-rejected examples without a separate reinforcement learning loop.
读寓言 →
Alignment
偏好数据
Preference Data
表达人类对多个候选输出偏好的数据，用于 RLHF、DPO 等对齐方法。
读寓言 →
Alignment
偏好优化
Preference Optimization
A class of methods that trains models to prefer better responses over worse ones using comparison data.
读寓言 →
Alignment
近端策略优化
Proximal Policy Optimization, PPO
The RL algorithm typically used in RLHF to update the language model policy while constraining updates to avoid destabilizing the model.
读寓言 →
Alignment
AI 反馈强化学习
Reinforcement Learning from AI Feedback, RLAIF
使用 AI 生成的偏好或批评信号来训练模型行为。
读寓言 →
Alignment
基于人类反馈的强化学习
Reinforcement Learning from Human Feedback, RLHF
An alignment technique that trains a reward model from human preferences and uses reinforcement learning to optimize model behavior.
读寓言 →
Alignment
奖励模型
Reward Model
A model trained to predict human preference scores, used in RLHF to assign rewards to model outputs during reinforcement learning.
读寓言 →
Application layer / Product
大模型应用
LLM Application, Large Model Application
大模型应用是把通用或行业大模型接入具体业务流程、数据和交互界面后形成的可用产品，价值重点通常在场景、数据、集成和交付，而不只是底层模型能力。
读寓言 →
Approximate nearest neighbor index
分层可导航小世界图
Hierarchical Navigable Small World, HNSW
HNSW is a graph-based ANN indexing algorithm widely used for fast vector similarity search.
读寓言 →
Architecture
注意力掩码
Attention Mask
A binary mask controlling which tokens can attend to which tokens; includes causal masks, padding masks, and custom patterns.
读寓言 →
Architecture
ALiBi
Attention with Linear Biases
A positional encoding method that adds a static, non-learned linear bias to attention scores that penalizes distant tokens.
读寓言 →
Architecture
BERT
Bidirectional Encoder Representations from Transformers
An encoder-only transformer pre-trained on masked language modeling and next-sentence prediction, designed for downstream understanding tasks.
读寓言 →
Architecture
因果注意力/自回归
Causal Attention / Autoregressive
Attention masked so each token only attends to itself and previous tokens, preserving temporal order for generation.
读寓言 →
Architecture
上下文长度/窗口
Context Window / Context Length
The maximum number of tokens a model can process in a single forward pass, a key capability and deployment constraint.
读寓言 →
Architecture
交叉注意力
Cross-Attention
Attention between two different sequences, typically connecting a decoder to encoder outputs in encoder-decoder models.
读寓言 →
Architecture
纯解码器
Decoder-Only
The architecture used by most modern LLMs (GPT family, Llama, Claude) where generation proceeds autoregressively without a separate encoder.
读寓言 →
Architecture
扩散语言模型
Diffusion Language Model
Non-autoregressive text generation using diffusion processes; an emerging alternative to autoregressive decoding for parallel generation.
读寓言 →
Architecture
编码器-解码器
Encoder-Decoder
A model architecture where an encoder processes the full input into a representation and a decoder generates output autoregressively from it.
读寓言 →
Architecture
前馈网络
Feed-Forward Network, FFN
The fully-connected MLP component in each transformer block that processes each token's representation independently after attention.
读寓言 →
Architecture
GPT
Generative Pre-trained Transformer
The line of decoder-only transformer models trained on next-token prediction and scaled to large sizes, originating from OpenAI.
读寓言 →
Architecture
分组查询注意力
Grouped Query Attention, GQA
An attention variant that shares key-value heads across groups of query heads, balancing speed (MQA) and quality (MHA).
读寓言 →
Architecture
Jamba
Jamba
A hybrid architecture combining Mamba SSM layers with transformer attention layers to leverage both linear and quadratic attention strengths.
读寓言 →
Architecture
层归一化
Layer Normalization
Normalizing activations within a layer to stabilize training; variants include Pre-LN (applied before attention/FFN) and Post-LN.
读寓言 →
Architecture
线性注意力
Linear Attention
Attention variants that reduce quadratic O(n²) complexity to O(n) by approximating the softmax using kernel methods.
读寓言 →
Architecture
Mamba
Mamba
A selective state space model with input-dependent parameters that achieves linear-time sequence processing while matching transformer quality.
读寓言 →
Architecture
MoD
Mixture of Depths
Dynamically routing tokens to be processed by a subset of transformer layers, skipping layers for tokens that don't need full depth.
读寓言 →
Architecture
混合专家模型
Mixture of Experts, MoE
A model architecture where different input tokens are routed to different expert sub-networks, enabling massive parameter counts with sub-linear compute per token.
读寓言 →
Architecture
多查询注意力
Multi-Query Attention, MQA
An attention variant where all query heads share a single key-value head, dramatically reducing KV cache size.
读寓言 →
Architecture
残差连接
Residual Connection
Skip connections that add a layer's input to its output, enabling stable gradient flow in very deep networks.
读寓言 →
Architecture
RMSNorm
Root Mean Square Layer Normalization
A simplified layer normalization that uses root mean square instead of standard deviation, reducing computation overhead.
读寓言 →
Architecture
旋转位置编码
Rotary Position Embedding, RoPE
A widely adopted positional encoding that applies rotation matrices to query/key vectors, encoding relative position naturally.
读寓言 →
Architecture
滑动窗口注意力
Sliding Window Attention
Attention restricted to a local window of tokens around each position, reducing memory usage and enabling very long sequence processing.
读寓言 →
Architecture
状态空间模型
State Space Model, SSM
An alternative sequence model architecture (e.g., Mamba, S4) that scales linearly with sequence length, challenging transformer dominance on long contexts.
读寓言 →
Architecture
SwiGLU
Swish-Gated Linear Unit
An activation variant combining Swish and gated linear units, widely used as the FFN activation in modern LLMs for improved quality.
读寓言 →
Architecture
T5
Text-to-Text Transfer Transformer
An encoder-decoder model that frames all NLP tasks as text-to-text problems, enabling a unified training and transfer approach.
读寓言 →
Architecture
视觉 Transformer
Vision Transformer, ViT
Applying transformer architecture directly to sequences of image patches, treating them like tokens; competitive with CNNs at sufficient scale.
读寓言 →
Architecture
权重绑定
Weight Tying
Sharing the weights of the embedding layer and the final output projection layer, reducing parameter count without quality loss.
读寓言 →
Auditability
审计证据
Audit Evidence
Records, logs, artifacts, screenshots, approvals, reports, or other proof used to support audit conclusions.
读寓言 →
Auditability
审计轨迹
Audit Trail
A chronological record of events showing who or what performed actions, when they occurred, and what changed.
读寓言 →
Auditability
可审计性
Auditability
The ability to inspect, reconstruct, and evidence system behavior, decisions, access, approvals, and control operation.
读寓言 →
Auditability
内部审计
Internal Audit
An independent assurance function that evaluates governance, risk management, and controls, including AI-related processes and systems.
读寓言 →
Auditability/Documentation
来源引用
Citation
A reference to source material used to support an AI output, decision, or generated claim.
读寓言 →
Auditability/Security
不可篡改日志
Tamper-Evident Log
A logging mechanism designed so unauthorized changes or deletions can be detected or prevented.
读寓言 →
Automation architecture
智能体工作流
Agentic Workflow
An agentic workflow combines repeatable process structure with agent-like autonomy, allowing models to choose tools, branch, revise steps, or recover from failures within bounded constraints.
读寓言 →
Automation architecture
工作流
Workflow
A workflow is a predefined or semi-structured sequence of steps, decisions, and integrations that coordinates model calls, tools, humans, and systems to produce a repeatable outcome.
读寓言 →
Backend engineering / operations
幂等性
Idempotency
幂等性是同一操作执行一次或多次都产生相同结果的性质，常用于重试、支付、任务调度和外部写入场景。
读寓言 →
Backend engineering / operations
任务队列
Job Queue, Task Queue
任务队列把耗时或异步工作排队执行，用于削峰、重试、调度和解耦前端请求与后台处理。
读寓言 →
Billing / platform
计量
Metering
计量是记录和汇总客户实际使用量的过程，是用量定价、成本归因和配额控制的基础。
读寓言 →
Business / SaaS metrics
年度经常性收入
Annual Recurring Revenue, ARR
ARR 是订阅或经常性合同在一年维度上的标准化收入，是 SaaS 业务规模和增长的核心指标。
读寓言 →
Business / SaaS metrics
流失率
Churn Rate
流失率是客户或收入在某段时间内停止续费、减少使用或离开的比例。
读寓言 →
Business / SaaS metrics
净收入留存
Net Revenue Retention, NRR
NRR 衡量现有客户收入在续费、扩容、降级和流失后的净变化，反映产品持续创造价值的能力。
读寓言 →
Business / customer success
客户健康分
Customer Health Score
客户健康分综合使用频率、价值达成、支持工单、关系状态和续费风险等信号评估客户状态。
读寓言 →
Business / finance
毛利
Gross Margin
毛利是收入扣除直接交付成本后的剩余比例，在 AI 产品中通常受模型调用、算力、人工交付和支持成本影响。
读寓言 →
Business / finance
单位经济模型
Unit Economics
单位经济模型分析每个客户、任务、请求或交易的收入、成本和利润，判断业务能否规模化盈利。
读寓言 →
Business / go-to-market
企业销售
Enterprise Sales
企业销售面向大客户和组织采购，通常涉及多角色决策、安全审查、试点、合同和较长销售周期。
读寓言 →
Business / go-to-market
理想客户画像
Ideal Customer Profile, ICP
ICP 描述最适合购买和成功使用产品的客户类型，包括行业、规模、痛点、预算、触发事件和采购能力。
读寓言 →
Business / go-to-market
产品主导增长
Product-Led Growth, PLG
产品主导增长通过免费试用、自助上手、产品内转化和用户传播推动获客、激活和扩张。
读寓言 →
Business / go-to-market
销售主导增长
Sales-Led Growth
销售主导增长依靠销售团队识别、推进和成交客户，常见于高客单价、复杂采购或企业级产品。
读寓言 →
Business / growth
客户获取成本
Customer Acquisition Cost, CAC
客户获取成本是获得一个新客户所需的销售、市场和相关投入成本。
读寓言 →
Business / growth
客户生命周期价值
Customer Lifetime Value, LTV
客户生命周期价值是一个客户在整个合作周期内预计贡献的净收入或毛利。
读寓言 →
Business / operations
服务级别协议
Service Level Agreement, SLA
SLA 是服务商与客户之间对可用性、响应时间、支持范围和违约补偿等内容的正式承诺。
读寓言 →
Business / operations
Token 成本
Token Cost
Token 成本是按模型输入和输出 token 数量计费或核算的使用成本，是 LLM 应用毛利和定价的重要变量。
读寓言 →
Business / post-sales
客户成功
Customer Success
客户成功通过培训、运营建议、价值追踪和续费扩容管理帮助客户持续获得结果。
读寓言 →
Business model / delivery
专业服务
Professional Services
专业服务是围绕咨询、实施、定制、培训和集成提供的人力密集型交付，常作为产品落地的补充。
读寓言 →
Business model / operations
托管服务
Managed Service
托管服务是供应商持续代客户运营某项系统或业务流程，通常包含软件、人工支持和结果责任。
读寓言 →
Business model / pricing
混合定价
Hybrid Pricing
混合定价同时结合订阅、席位、用量、服务费或结果分成，以匹配不同成本结构和客户价值感知。
读寓言 →
Business model / pricing
结果定价
Outcome-Based Pricing
结果定价按照客户获得的业务结果、节省成本或新增收入收费，而不是只按席位、时间或调用量收费。
读寓言 →
Business model / pricing
席位定价
Seat-Based Pricing, Per-Seat Pricing
席位定价按使用产品的人数或账号数量收费，适合人与软件交互频繁且价值与用户规模相关的产品。
读寓言 →
Business model / pricing
用量定价
Usage-Based Pricing
用量定价根据客户实际使用量收费，例如请求数、token 数、任务数、存储量或处理数据量。
读寓言 →
Business model / product
软件即服务
SaaS, Software as a Service
SaaS 是通过在线软件持续交付标准化产品能力，并通常按订阅、席位或用量收费的商业模式。
读寓言 →
Business model / product strategy
产品化服务
Productized Service
产品化服务是把定制服务压缩成清晰范围、固定流程、明确价格和可重复交付的标准化方案。
读寓言 →
Business positioning / Automation
数字员工
Digital Employee, Digital Worker, AI Worker
数字员工是把 AI 能力包装成可承担岗位任务或流程结果的虚拟劳动力，市场表达上更接近“替代或增强某类工作”而不是单纯卖软件功能。
读寓言 →
Capability packaging
技能
Skill
A skill is a packaged capability that gives an agent task-specific instructions, tools, examples, or procedures for reliably performing a class of work.
读寓言 →
Compliance
合规映射
Compliance Mapping
The process of linking policies, controls, evidence, and system capabilities to specific regulatory, contractual, or framework requirements.
读寓言 →
Compliance/Audit
控制测试
Control Testing
Evidence-based verification that a control is designed properly and operating effectively.
读寓言 →
Compliance/Documentation
技术文档
Technical Documentation
Required documentation describing an AI system's purpose, design, data, model behavior, risk controls, evaluation, monitoring, and operating instructions.
读寓言 →
Compliance/Governance
控制框架
Control Framework
A structured set of control objectives and activities used to manage risk and demonstrate compliance across domains.
读寓言 →
Compliance/Security
控制措施
Controls
Policies, procedures, technical safeguards, or operational checks designed to reduce risk or enforce requirements.
读寓言 →
Compliance/Security
ISO 27001
ISO/IEC 27001
An international standard for establishing, operating, monitoring, and improving an information security management system.
读寓言 →
Compliance/Security
SOC 2 Type I
SOC 2 Type I
A SOC 2 report assessing whether controls are suitably designed at a specific point in time.
读寓言 →
Compliance/Security
SOC 2 Type II
SOC 2 Type II
A SOC 2 report assessing the operating effectiveness of controls over a defined review period.
读寓言 →
Compliance/Security
SOC 2
Service Organization Control 2, SOC 2
An assurance report evaluating a service organization's controls against trust services criteria such as security, availability, confidentiality, processing integrity, and privacy.
读寓言 →
Compression
AWQ
Activation-Aware Weight Quantization
A quantization method that protects salient weight channels (identified by activation magnitudes) while aggressively quantizing less important channels.
读寓言 →
Compression
bitsandbytes
BitsAndBytes
A Python library for low-precision model loading (4-bit, 8-bit) and quantization, commonly used with Hugging Face Transformers.
读寓言 →
Compression
深度剪枝
Depth Pruning
Removing entire transformer layers from a model, preserving output quality through a shallower architecture.
读寓言 →
Compression
蒸馏推理模型
Distilled Reasoning Models
Using outputs from a large reasoning model (like DeepSeek-R1) to fine-tune a smaller model, transferring reasoning capabilities.
读寓言 →
Compression
GGML
GGML Tensor Library
The original C library behind llama.cpp for CPU-based LLM inference with quantization support; precursor to GGUF.
读寓言 →
Compression
GPTQ
GPT Post-Training Quantization
A one-shot weight quantization method using approximate second-order information (Hessian) to minimize 3/4-bit quantization error.
读寓言 →
Compression
GGUF
GPT-Generated Unified Format
A file format for quantized LLMs enabling CPU inference via llama.cpp; supports various quantization levels (q4_0, q5_K_M, etc.).
读寓言 →
Compression
模型剪枝
Model Pruning
Removing less important weights, neurons, or attention heads from a trained model to reduce size and latency while preserving quality.
读寓言 →
Compression
训练后量化
Post-Training Quantization, PTQ
Quantizing model weights after training without further fine-tuning; fast but potentially lossy.
读寓言 →
Compression
量化感知训练
Quantization-Aware Training, QAT
Training or fine-tuning with simulated quantization effects so the model adapts to lower precision.
读寓言 →
Compression
稀疏性
Sparsity
The property of having many zero-valued weights or activations; can be induced via pruning for computational savings.
读寓言 →
Compression
结构化剪枝
Structured Pruning
Pruning entire neurons, heads, or layers rather than individual weights, producing models that run efficiently on standard hardware.
读寓言 →
Compression
非结构化剪枝
Unstructured Pruning
Zeroing individual weights based on importance, producing sparse weight matrices that may need special hardware/runtimes to benefit.
读寓言 →
Compression
宽度剪枝
Width Pruning
Reducing hidden dimensions and intermediate sizes of FFN or attention layers for a thinner but equally deep model.
读寓言 →
Concept distinction
思维链与草稿区
Chain-of-Thought vs Scratchpad
Chain-of-thought is a reasoning process or trace, while a scratchpad is a broader temporary state area that can include reasoning, observations, tool outputs, intermediate data, and execution notes.
读寓言 →
Content provenance
C2PA
Coalition for Content Provenance and Authenticity
A standard framework for certifying digital content origin and history.
读寓言 →
Control pattern
精修版人在回路
Human-in-the-Loop
Human-in-the-loop design inserts human review, approval, correction, or escalation into an agent process when judgment, risk, compliance, or irreversible action requires oversight.
读寓言 →
Conversation state object
会话线程
Thread
A thread stores the ongoing conversation context and messages between a user and an assistant in the Assistants API model.
读寓言 →
Coordination pattern
交接
Handoff
Handoff is the transfer of task ownership, context, and next actions from one agent, workflow, system, or human to another.
读寓言 →
Cost / context
Token 预算
Token Budget
为任务、请求、用户或工作流分配的最大 token 使用额度。
读寓言 →
Cost optimization
预算感知路由
Budget-Aware Routing
按成本预算、质量和延迟选择模型、工具或执行路径。
读寓言 →
Customer support / Contact center AI
智能客服
AI Customer Service, Intelligent Customer Service
智能客服用自然语言理解、知识库检索、对话管理和工单系统集成来自动响应客户咨询、分流问题或辅助人工客服。
读寓言 →
Data framework for LLM applications
LlamaIndex
LlamaIndex
LlamaIndex is a data and retrieval framework that helps connect LLM applications to private, structured, semi-structured, and unstructured data sources.
读寓言 →
Data ingestion
文档解析
Document Parsing
把 PDF、Word、网页、表格等转成可索引文本和结构化元素。
读寓言 →
Data ingestion
OCR
Optical Character Recognition
从图片或扫描件中识别文字。
读寓言 →
Data ingestion integration
数据连接器
Data Connector
A data connector loads external files, databases, SaaS content, webpages, or APIs into an AI application's indexing and retrieval pipeline.
读寓言 →
Data strategy / Competitive advantage
数据飞轮
Data Flywheel
数据飞轮指产品使用、业务反馈和结果数据持续回流，反过来改进模型、流程、推荐、评估或交付质量，从而形成越用越好的增长机制。
读寓言 →
Decoding
束搜索
Beam Search
A decoding strategy that keeps several high-probability candidate sequences at each step and selects among them.
读寓言 →
Decoding
解码
Decoding
The process of choosing output tokens from a model's predicted probability distribution during generation.
读寓言 →
Decoding
贪心解码
Greedy Decoding
A decoding strategy that always selects the single most likely next token.
读寓言 →
Decoding
Logits
Logits
The raw scores a model outputs before they are converted into probabilities over possible next tokens.
读寓言 →
Decoding
采样
Sampling
A decoding strategy that randomly selects tokens according to the model's probability distribution, often to increase diversity.
读寓言 →
Decoding
温度
Temperature
A generation parameter that controls randomness by flattening or sharpening the next-token probability distribution.
读寓言 →
Decoding
Top-k 采样
Top-k Sampling
A sampling method that restricts next-token choices to the k most likely tokens.
读寓言 →
Decoding
Top-p 采样
Top-p Sampling, Nucleus Sampling
A sampling method that restricts next-token choices to the smallest set of tokens whose cumulative probability exceeds p.
读寓言 →
Deep Learning
反向传播
Backpropagation
The algorithm used to compute gradients through a neural network so its parameters can be updated during training.
读寓言 →
Deep Learning
深度学习
Deep Learning
A branch of machine learning that uses multi-layer neural networks to learn complex representations from large amounts of data.
读寓言 →
Deep Learning
神经网络
Neural Network
A model architecture composed of layers of connected units that learn nonlinear transformations from data.
读寓言 →
Deep Learning
表征学习
Representation Learning
The process by which models learn useful internal features or representations of raw data for downstream tasks.
读寓言 →
Deployment
金丝雀发布
Canary Release
先向少量用户或流量发布新版本，观察稳定后再扩大范围。
读寓言 →
Deployment
灰度发布
Gradual Rollout
按用户、租户、区域或流量比例逐步扩大新版本覆盖。
读寓言 →
Deployment / Enterprise procurement
私有化部署
Private Deployment, On-premises Deployment
私有化部署通常指 AI 系统部署在客户自有或专属控制的环境中，以满足数据安全、合规、内网访问、权限管理和定制集成要求。
读寓言 →
Deployment / Infrastructure
本地化部署
Local Deployment, On-premises Deployment
本地化部署强调模型、应用或推理服务在客户本地机房、内网服务器、边缘设备或专属算力环境中运行，常被用于降低外部数据传输和网络依赖。
读寓言 →
Deployment / infrastructure
容器
Container
容器把应用及其依赖打包成可移植运行单元，使部署环境更一致并便于隔离和扩缩容。
读寓言 →
Deployment / infrastructure
Kubernetes
Kubernetes, K8s
Kubernetes 是用于编排容器化应用的系统，负责调度、扩缩容、服务发现、滚动发布和自愈。
读寓言 →
Deployment / release engineering
蓝绿部署
Blue-Green Deployment
蓝绿部署通过维护两套生产环境并切换流量来降低发布风险和加快回滚。
读寓言 →
Deployment / release engineering
金丝雀发布
Canary Deployment
金丝雀发布先向少量用户或流量推出新版本，观察指标稳定后再逐步扩大范围。
读寓言 →
DevOps / infrastructure
基础设施即代码
Infrastructure as Code, IaC
基础设施即代码用声明式或脚本化配置管理服务器、网络、权限和云资源，使环境可审查、可复现、可版本化。
读寓言 →
DevOps / operations
GitOps
GitOps
GitOps 以 Git 仓库作为期望状态来源，通过自动化控制器把基础设施和应用环境同步到声明配置。
读寓言 →
DevOps / release engineering
持续集成与持续交付
CI/CD, Continuous Integration / Continuous Delivery
CI/CD 是通过自动化构建、测试和发布流水线让代码更频繁、可靠地交付到生产环境。
读寓言 →
Dialogue / workflow
插槽填充
Slot Filling
从用户输入中抽取流程所需字段，如时间、地点、金额、客户名。
读寓言 →
Diffusion
噪声预测器/UNet
UNet / Noise Predictor
The backbone network in classic diffusion models that predicts the noise added at each step; typically a UNet with cross-attention for conditioning.
读寓言 →
Distributed inference
流水线并行
Pipeline Parallelism
Splitting model layers across devices so micro-batches flow through in a pipeline, reducing idle time through scheduling strategies.
读寓言 →
Distributed inference
张量并行
Tensor Parallelism
Splitting individual weight matrices across multiple GPUs so each GPU computes a portion of layer operations, enabling large models to fit across devices.
读寓言 →
Distributed training
数据并行
Data Parallelism
Replicating the model across devices, each processing a different data batch, with synchronized gradient updates.
读寓言 →
Distributed training
FSDP
Fully Sharded Data Parallelism
A data parallel strategy that shards model parameters, gradients, and optimizer states across GPUs to reduce per-device memory.
读寓言 →
Distributed training
ZeRO
Zero Redundancy Optimizer
DeepSpeed's family of memory optimization stages (1, 2, 3) that partition optimizer states, gradients, and parameters across GPUs.
读寓言 →
Document AI
版面分析
Layout Analysis
识别文档中的标题、段落、表格、图片和阅读顺序。
读寓言 →
Document AI
表格抽取
Table Extraction
从文档或图片中识别表格结构与单元格内容。
读寓言 →
Documentation
数据集说明书
Datasheet for Datasets
A structured description of dataset motivation, composition, collection, preprocessing, uses, and risks.
读寓言 →
Documentation
模型卡
Model Card
A structured document describing model purpose, training, evaluation, limitations, and ethical considerations.
读寓言 →
Documentation
系统卡
System Card
A document describing a deployed AI system, its capabilities, risks, mitigations, and evaluation results.
读寓言 →
Embedding database
Chroma
Chroma
Chroma is an open-source embedding database often used by developers for local or lightweight RAG prototypes and applications.
读寓言 →
Embeddings
嵌入
Embedding
A dense numerical vector that represents text, images, audio, users, or other objects in a way that captures useful semantic relationships.
读寓言 →
Embeddings
语义相似度
Semantic Similarity
A measure of how close two pieces of content are in meaning, often computed using embedding vectors.
读寓言 →
Embeddings
向量数据库
Vector Database
A database optimized for storing embedding vectors and retrieving nearest neighbors by similarity.
读寓言 →
Embeddings
词嵌入
Word Embedding
A vector representation of a word or token that captures statistical and semantic relationships learned from text.
读寓言 →
Engineering organization / infrastructure
平台工程
Platform Engineering
平台工程通过内部开发者平台和自助能力提升工程团队交付效率、可靠性和治理一致性。
读寓言 →
Engineering organization / platform
开发者体验
Developer Experience, DevEx
开发者体验衡量工程师从理解、开发、测试、部署到排障的效率、顺畅度和认知负担。
读寓言 →
Enterprise AI application / Retrieval
知识库问答
Knowledge Base Q&A, KBQA, RAG Q&A
知识库问答是将企业文档、制度、产品资料或业务知识接入检索与生成系统，让用户用自然语言获得带上下文依据的答案。
读寓言 →
Enterprise AI orchestration SDK
Semantic Kernel
Semantic Kernel
Semantic Kernel is Microsoft's SDK for integrating AI models with plugins, planners, memory, prompts, and enterprise application code.
读寓言 →
Enterprise collaboration / Solution design
共创
Co-creation, Joint Solution Development
共创是供应商与客户围绕真实业务问题共同定义场景、数据、流程、指标和方案的合作方式，常用于早期行业方案打磨或关键客户深度绑定。
读寓言 →
Enterprise implementation / Adoption
试点
Pilot, Pilot Program
试点是在真实业务环境中小范围运行 AI 方案，以验证用户采纳、流程适配、运营成本、稳定性和可扩展复制条件。
读寓言 →
Evaluation
评测基准
Benchmark
A standardized task, dataset, or suite used to compare model capabilities and limitations.
读寓言 →
Evaluation
基准污染
Benchmark Contamination
A form of data leakage where benchmark questions or answers appear in training data, inflating measured model performance.
读寓言 →
Evaluation
数据泄漏
Data Leakage
A problem where information from evaluation or target data improperly appears in training or model selection, making performance estimates misleading.
读寓言 →
Evaluation
分布偏移
Distribution Shift
A mismatch between the data distribution a model was trained on and the data it encounters in deployment.
读寓言 →
Evaluation
困惑度
Perplexity
A language modeling metric that measures how well a model predicts a sequence, with lower values indicating better predictive fit.
读寓言 →
Evaluation
步骤成功率
Step Success Rate
衡量多步骤任务中单步执行正确的比例。
读寓言 →
Evaluation
工具调用准确率
Tool Call Accuracy
衡量 Agent 是否选择了正确工具并传入正确参数。
读寓言 →
Evaluation
轨迹评估
Trajectory Evaluation
对 Agent 的中间步骤、工具调用、观察和最终结果进行整体评估。
读寓言 →
Evaluation infrastructure
LLM 评测平台
LLM Evaluation Platform
An LLM evaluation platform manages datasets, test cases, scoring functions, regression checks, and experiment comparison for model-powered applications.
读寓言 →
FinOps / operations
成本归因
Cost Attribution
成本归因把云资源、模型调用和人工成本分摊到客户、租户、功能或任务上，帮助判断利润和优化定价。
读寓言 →
Fine-Tuning
微调
Fine-Tuning
Additional training of a pretrained model on narrower data to improve performance for specific tasks, domains, or behaviors.
读寓言 →
Fine-Tuning
指令微调
Instruction Tuning
Training a model on instruction-response examples so it better follows natural language requests.
读寓言 →
Fine-Tuning
低秩适配
Low-Rank Adaptation, LoRA
A PEFT method that injects small trainable low-rank matrices into a model while keeping most original weights frozen.
读寓言 →
Fine-Tuning
参数高效微调
Parameter-Efficient Fine-Tuning, PEFT
A family of methods that adapt large models by training only a small number of additional or selected parameters.
读寓言 →
Fine-Tuning
监督微调
Supervised Fine-Tuning, SFT
Fine-tuning a pretrained model on curated input-output examples to teach task behavior or instruction following.
读寓言 →
Fine-tuning
全量微调
Full Fine-Tuning
Fine-tuning all model parameters on task-specific data, in contrast to parameter-efficient methods that freeze most weights.
读寓言 →
Fine-tuning
指令数据
Instruction Data
用于教模型遵循人类指令的数据，通常包含任务、输入和期望输出。
读寓言 →
Fine-tuning
分层学习率
Layer-wise Learning Rate
Applying different learning rates to different model layers during fine-tuning, often lower for early layers and higher for task-specific heads.
读寓言 →
Fine-tuning
拒绝采样
Rejection Sampling
Generating multiple candidate outputs and keeping only those that meet quality criteria to create high-quality training data for further fine-tuning.
读寓言 →
Foundations
人工智能
Artificial Intelligence, AI
A broad field focused on building systems that can perform tasks normally associated with human intelligence, such as perception, reasoning, language, planning, and decision-making.
读寓言 →
Foundations
模型
Model
A learned mathematical system that maps inputs to outputs based on patterns captured during training.
读寓言 →
Generative AI
扩散模型
Diffusion Model
A generative model that learns to create data by reversing a gradual noise-adding process.
读寓言 →
Generative modeling
扩散变换器
Diffusion Transformer, DiT
A transformer-based backbone for diffusion models that replaces UNet, enabling better scaling to high resolution and complex distributions.
读寓言 →
Generative modeling
潜在扩散模型
Latent Diffusion Model, LDM
A diffusion model that operates in a compressed latent space (via VAE) rather than pixel space, dramatically improving efficiency.
读寓言 →
Go-to-market / Implementation
场景化落地
Scenario-based Implementation, Use-case Implementation
场景化落地是从具体业务场景、角色任务、输入输出、系统集成和验收指标出发实施 AI，而不是泛泛提供通用模型能力。
读寓言 →
Governance
AI 影响评估
AI Impact Assessment, AIIA
A structured assessment of how an AI system may affect people, rights, safety, operations, compliance, and business outcomes.
读寓言 →
Governance
AI 使用场景清单
AI Use Case Inventory
A maintained catalog of AI use cases, owners, business purposes, data inputs, vendors, risk classifications, and operational status.
读寓言 →
Governance
数据血缘
Data Lineage
Tracking where data came from, how it was transformed, and where it was used.
读寓言 →
Governance
治理、风险与合规
Governance, Risk, and Compliance, GRC
The integrated discipline of managing organizational governance, risk processes, controls, policies, audits, and compliance obligations.
读寓言 →
Governance
策略引擎
Policy Engine
A component that evaluates rules to allow, deny, or constrain actions.
读寓言 →
Governance
来源证明
Provenance
Evidence about origin, authorship, or transformation history of data or content.
读寓言 →
Governance
影子 AI
Shadow AI
The use of AI tools, models, or agents without formal approval, security review, procurement visibility, or governance controls.
读寓言 →
Governance
可追溯性
Traceability
The ability to reconstruct why a system produced an output or took an action.
读寓言 →
Governance
透明度
Transparency
Clear disclosure of AI system purpose, capabilities, limitations, data use, decision role, and user or operator responsibilities.
读寓言 →
Governance
可信 AI
Trustworthy AI
AI that can be relied upon because it is lawful, ethical, technically robust, secure, explainable where needed, and subject to appropriate oversight.
读寓言 →
Governance and control
智能体策略
Agent Policy
An agent policy defines what the agent is allowed, required, or forbidden to do across tools, data, users, external actions, and escalation paths.
读寓言 →
Governance/Compliance
AI 披露
AI Disclosure
Informing users, customers, regulators, or affected individuals when AI is used, what role it plays, and what rights or options they have.
读寓言 →
Governance/Compliance
控制所有人
Control Owner
The person or function accountable for designing, operating, evidencing, and improving a specific control.
读寓言 →
Governance/Operations
人工监督
Human Oversight
Human review, intervention, escalation, or override mechanisms used to prevent or mitigate harmful or inappropriate AI outcomes.
读寓言 →
Governance/Risk
偏见
Bias
Systematic skew in data, model behavior, design, or deployment that can produce inaccurate, unfair, or harmful outcomes.
读寓言 →
Governance/Risk
可解释性
Explainability
The ability to provide understandable reasons, factors, or evidence for an AI system's output or decision at the level required by the use case.
读寓言 →
Governance/Risk
公平性
Fairness
The property that an AI system avoids unjustified adverse treatment or outcomes across protected or relevant groups.
读寓言 →
Governance/Risk
三道防线
Three Lines of Defense
A governance model separating business ownership, risk/compliance oversight, and independent audit responsibilities.
读寓言 →
Governance/Safety
内容过滤
Content Filtering
Screening AI inputs or outputs for prohibited, unsafe, sensitive, or policy-violating content before further use.
读寓言 →
Governance/Security
策略即代码
Policy as Code
The practice of expressing governance, security, or compliance rules in machine-enforceable code that can be tested, versioned, and audited.
读寓言 →
IT operations / automation
人工智能运维
AIOps
AIOps 是用机器学习和自动化分析运维数据，辅助异常检测、告警降噪、根因分析和事件响应。
读寓言 →
Implementation / Project governance
交付验收
Delivery Acceptance, Acceptance Testing, Project Acceptance
交付验收是客户依据合同、需求文档、功能清单、性能指标、安全要求和业务效果标准确认 AI 项目达到可交付状态的过程。
读寓言 →
Inference
自回归解码
Autoregressive Decoding / Decode Phase
The sequential phase where tokens are generated one at a time, each step attending to the full KV cache.
读寓言 →
Inference
最佳-of-N 采样
Best-of-N Sampling
Generating N candidate outputs and selecting the best one via a reward model or verifier; a simple form of test-time compute scaling.
读寓言 →
Inference
上下文剪枝
Context Pruning / Prompt Compression
Reducing prompt length by removing or compressing less relevant tokens to reduce KV cache size and prefill latency.
读寓言 →
Inference
动态批处理
Dynamic Batching / In-flight Batching
Adding new requests to an actively processing batch as earlier requests finish, rather than waiting for the entire batch to complete.
读寓言 →
Inference
早出
Early Exiting
Stopping computation at an intermediate layer when confidence is high enough, reducing latency for easy queries.
读寓言 →
Inference
Flash Attention
FlashAttention
An IO-aware attention algorithm that minimizes HBM reads/writes by fusing operations and using tiling, reducing memory footprint from O(n²) to sub-quadratic.
读寓言 →
Inference
推理
Inference
The process of running a trained model to produce outputs from new inputs.
读寓言 →
Inference
缓存卸载
KV Cache Offloading
Moving KV cache to CPU RAM or SSD to handle very long contexts on limited GPU memory, at the cost of increased latency.
读寓言 →
Inference
量化 KV 缓存
KV Cache Quantization
Storing KV cache entries at lower precision (8-bit, 4-bit) to support longer contexts within limited GPU memory.
读寓言 →
Inference
算子融合
Kernel Fusion
Combining multiple GPU operations into a single kernel to reduce memory round-trips and launch overhead; used in Flash Attention and optimized runtimes.
读寓言 →
Inference
延迟
Latency
The time it takes for a model system to respond to a request or generate tokens.
读寓言 →
Inference
回顾解码
Lookahead Decoding
Generating n-gram candidate sequences from Jacobi iteration and verifying them in parallel without a draft model.
读寓言 →
Inference
Medusa
Medusa
An extension of speculative decoding that adds multiple decoding heads to the base model, predicting several future tokens in parallel without a separate draft model.
读寓言 →
Inference
模型服务
Model Serving
The production infrastructure and APIs used to host models and respond to inference requests.
读寓言 →
Inference
分页注意力
Paged Attention
A memory management technique that stores KV cache in non-contiguous pages, eliminating fragmentation and enabling near-optimal memory utilization; powers vLLM.
读寓言 →
Inference
预填充
Prefill Phase
The initial phase of LLM inference where all input tokens are processed in parallel to populate the KV cache before step-by-step decoding begins.
读寓言 →
Inference
前缀缓存
Prefix Caching
Reusing KV cache entries from shared prompt prefixes across requests (e.g., system prompts), avoiding redundant prefill computation.
读寓言 →
Inference
量化
Quantization
A compression and acceleration technique that represents model weights or activations with lower-precision numbers.
读寓言 →
Inference
推测解码
Speculative Decoding
Using a fast draft model to generate candidate tokens cheaply, which a larger target model then verifies in parallel, accelerating generation 2-3× without quality loss.
读寓言 →
Inference
推测编辑
Speculative Editing
A variant of speculative decoding focused on assistant drafting with editing, used in some production serving systems.
读寓言 →
Inference
推测前缀树
Speculative Prefix Tree
A tree-structured approach to speculative decoding where multiple candidate paths are explored and verified in parallel.
读寓言 →
Inference
投机采样
Speculative Sampling
Alternate name for speculative decoding; generating draft tokens speculatively and accepting/rejecting via verification.
读寓言 →
Inference
测试时计算
Test-Time Compute / Inference-Time Scaling
Allocating additional compute at inference time (e.g., longer chain-of-thought, best-of-N, tree search) to improve output quality without model changes.
读寓言 →
Inference
吞吐量
Throughput
The amount of work a model serving system can process over time, often measured in requests or tokens per second.
读寓言 →
Inference
树搜索
Tree Search
Exploring multiple reasoning paths in parallel via Monte Carlo tree search (MCTS) or beam search for tasks like math and code.
读寓言 →
Inference optimization
连续批处理
Continuous Batching
连续批处理是在推理过程中动态合并和调度不同请求，以提高 GPU 利用率和吞吐量。
读寓言 →
Inference optimization
提示缓存
Prompt Caching
Prompt caching reuses previously processed prompt prefixes or repeated context to reduce latency and cost when many requests share the same instructions or reference material.
读寓言 →
Infrastructure
运行时
Runtime
The environment that executes an agent, tools, permissions, memory, and lifecycle control.
读寓言 →
Knowledge access
检索增强生成
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation supplies a model with relevant external documents or records at inference time so responses can be grounded in up-to-date or private knowledge.
读寓言 →
Knowledge representation
知识图谱
Knowledge Graph
A knowledge graph represents entities and their relationships as structured nodes and edges so machines can query, reason over, and connect facts explicitly.
读寓言 →
Knowledge representation
本体
Ontology
定义领域概念、属性和关系的结构化知识模型。
读寓言 →
Knowledge representation
三元组
Triple
主语-谓语-宾语形式的知识表示单元。
读寓言 →
Knowledge-augmented generation
图增强检索生成
Graph RAG
Graph RAG uses graph structures, entity relationships, and graph traversal or community summaries to retrieve context that simple chunk-level vector search may miss.
读寓言 →
LLM
上下文窗口
Context Window
The maximum amount of input and generated text, measured in tokens, that a model can attend to in a single request.
读寓言 →
LLM
语言模型
Language Model
A model that assigns probabilities to sequences of language and can predict likely next tokens.
读寓言 →
LLM
大语言模型
Large Language Model, LLM
A large neural language model, usually transformer-based, trained on massive text corpora to predict and generate language.
读寓言 →
LLM API gateway and SDK
LiteLLM
LiteLLM
LiteLLM provides a unified OpenAI-style interface, proxy, routing, cost tracking, and policy layer across many model providers.
读寓言 →
LLM Agents
智能体
Agent
A system that uses a model to plan, call tools, maintain state, and act toward a goal across one or more steps.
读寓言 →
LLM Agents
记忆
Memory
Stored information that lets an AI system preserve useful context across steps, sessions, or user interactions.
读寓言 →
LLM Agents
规划
Planning
The process by which an AI system decomposes a goal into steps, chooses actions, and adapts as conditions change.
读寓言 →
LLM Agents
工具调用
Tool Use, Function Calling
A capability where a model requests external tools or APIs to retrieve information, compute results, or take actions.
读寓言 →
LLM Applications
检索增强生成
Retrieval-Augmented Generation, RAG
A pattern where a model retrieves relevant external information and uses it as context when generating an answer.
读寓言 →
LLM Behavior
精修版幻觉
Hallucination
A failure mode where a model generates plausible-sounding but false, unsupported, or fabricated information.
读寓言 →
LLM Interaction
少样本提示
Few-Shot Prompting
Providing a small number of examples in the prompt to demonstrate the desired task or output format.
读寓言 →
LLM Interaction
上下文学习
In-Context Learning
A model's ability to adapt behavior based on examples or instructions in the prompt without updating its weights.
读寓言 →
LLM Interaction
提示词
Prompt
The input instructions, context, examples, and constraints given to a model to guide its output.
读寓言 →
LLM Interaction
提示工程
Prompt Engineering
The practice of designing prompts to steer model behavior and improve task performance without changing model weights.
读寓言 →
LLM Interaction
系统提示词
System Prompt, System Message
A high-priority instruction that defines model behavior, role, constraints, and policy within a conversation or application.
读寓言 →
LLM Interaction
零样本提示
Zero-Shot Prompting
Asking a model to perform a task without providing examples in the prompt.
读寓言 →
LLM Reasoning
思维链
Chain-of-Thought, CoT
A prompting or training pattern where intermediate reasoning steps are used to help solve complex tasks.
读寓言 →
LLM application framework
LangChain
LangChain
LangChain is a framework for building LLM applications by composing prompts, models, retrievers, tools, chains, and agents into reusable workflows.
读寓言 →
LLM evaluation and observability library
TruLens
TruLens
TruLens helps instrument and evaluate LLM applications with feedback functions for relevance, groundedness, toxicity, and other quality signals.
读寓言 →
LLM inference
解码阶段
Decode Phase
自回归逐 token 生成输出的阶段。
读寓言 →
LLM inference
预填充
Prefill
推理时处理输入上下文并构建初始 KV cache 的阶段。
读寓言 →
LLM inference optimization
分页注意力
PagedAttention
PagedAttention is a memory-management technique that stores key-value cache blocks efficiently to improve throughput for LLM serving.
读寓言 →
LLM inference server
文本生成推理
Text Generation Inference, TGI
TGI is Hugging Face's production-oriented inference server for deploying and serving text generation models with batching, streaming, and monitoring features.
读寓言 →
LLM infrastructure optimization
缓存
Caching
Caching stores reusable responses, embeddings, retrieval results, or prompt prefixes to reduce latency, cost, and repeated model calls.
读寓言 →
LLM observability and evaluation platform
LangSmith
LangSmith
LangSmith is LangChain's platform for tracing, debugging, evaluating, monitoring, and testing LLM applications and agent workflows.
读寓言 →
LLM observability and gateway tool
Helicone
Helicone
Helicone provides logging, monitoring, caching, cost tracking, and analytics for LLM API usage, often through a proxy-style integration.
读寓言 →
LLM operations
成本跟踪
Cost Tracking
Cost tracking measures token usage, provider charges, cache savings, and per-user or per-workflow spend so teams can manage unit economics.
读寓言 →
LLM orchestration abstraction
LangChain 表达式语言
LangChain Expression Language, LCEL
LCEL is LangChain's declarative composition layer for connecting prompts, models, parsers, retrievers, and custom functions into runnable pipelines.
读寓言 →
LLM program interface
签名
Signature
In DSPy, a signature declares the input and output fields of an LLM task so the system can compile and optimize the prompting strategy.
读寓言 →
LLM programming and optimization framework
DSPy
DSPy
DSPy is a framework for programming LLM systems with declarative modules and optimizing prompts or demonstrations against metrics.
读寓言 →
LLM runtime / optimization
KV 缓存
KV Cache, Key-Value Cache
KV 缓存保存 Transformer 已计算的注意力键和值，避免生成每个新 token 时重复计算历史上下文。
读寓言 →
LLM runtime / user experience
首 Token 时间
Time to First Token, TTFT
首 Token 时间是流式生成中从请求发出到收到第一个 token 的时间，是感知响应速度的关键指标。
读寓言 →
LLMOps / application platform
提示词管理
Prompt Management
提示词管理是对提示词模板、变量、版本、实验、审批和效果追踪进行系统化管理的实践。
读寓言 →
LLMOps / quality
评测
Evaluation, Evals
评测是用人工、规则或模型裁判衡量 AI 系统在任务质量、安全性、稳定性和成本上的表现。
读寓言 →
LLMOps / quality
黄金数据集
Golden Dataset, Golden Set
黄金数据集是一组经过人工确认、可重复使用的高质量样例，用于回归评测和模型或提示词对比。
读寓言 →
Local model packaging
模型文件配置
Modelfile
A Modelfile defines how an Ollama model is built or customized, including base model, parameters, system prompt, and template settings.
读寓言 →
Local model runtime
Ollama
Ollama
Ollama is a local runtime and model management tool that makes it easy to run, pull, package, and serve LLMs on personal or private machines.
读寓言 →
MLOps / data platform
特征库
Feature Store
特征库是集中生产、存储、复用和服务机器学习特征的数据平台组件。
读寓言 →
MLOps / governance
模型版本管理
Model Versioning
模型版本管理用于追踪不同模型、数据、参数和代码组合，确保部署、回滚和审计可复现。
读寓言 →
MLOps / monitoring
模型漂移
Model Drift
模型漂移是生产环境中的输入分布、目标关系或业务语境变化导致模型效果下降的现象。
读寓言 →
MLOps / platform
模型注册表
Model Registry
模型注册表是集中管理模型版本、元数据、审批状态和部署流转的系统。
读寓言 →
Machine Learning
泛化
Generalization
A model's ability to perform well on new examples that were not seen during training.
读寓言 →
Machine Learning
梯度下降
Gradient Descent
An optimization method that updates model parameters in the direction that reduces the loss function.
读寓言 →
Machine Learning
超参数
Hyperparameters
Configuration choices set before or during training, such as learning rate, batch size, and number of layers, that control how a model learns.
读寓言 →
Machine Learning
损失函数
Loss Function
A mathematical objective that measures how wrong a model's output is and guides parameter updates during training.
读寓言 →
Machine Learning
机器学习
Machine Learning, ML
A subfield of AI where systems learn patterns from data to make predictions or decisions without being explicitly programmed for every rule.
读寓言 →
Machine Learning
过拟合
Overfitting
A failure mode where a model performs well on training data but poorly on new data because it memorized noise or overly specific patterns.
读寓言 →
Machine Learning
参数
Parameters
The learned numerical values inside a model that determine how it transforms inputs into outputs.
读寓言 →
Machine Learning
强化学习
Reinforcement Learning, RL
A learning paradigm where an agent learns actions by receiving rewards or penalties from interaction with an environment.
读寓言 →
Machine Learning
监督学习
Supervised Learning
A learning setup where a model is trained on examples that include both inputs and known target labels or outputs.
读寓言 →
Machine Learning
测试集
Test Set
A held-out dataset used to estimate final model performance on unseen data after development choices are fixed.
读寓言 →
Machine Learning
训练集
Training Set
The data used to fit a model's parameters during training.
读寓言 →
Machine Learning
无监督学习
Unsupervised Learning
A learning setup where a model finds structure in data without explicit target labels.
读寓言 →
Machine Learning
验证集
Validation Set
A held-out dataset used during development to tune model choices and estimate generalization before final testing.
读寓言 →
Managed agent API
OpenAI 助手 API
OpenAI Assistants API
The Assistants API is OpenAI's managed API for building assistants with threads, messages, runs, tools, files, and persistent conversation state.
读寓言 →
Managed agent object
助手
Assistant
An assistant is a configured AI entity with instructions, model choice, tools, and optional files or resources for handling user tasks.
读寓言 →
Managed model and agent API
OpenAI 响应 API
OpenAI Responses API
The Responses API is OpenAI's unified interface for model responses, tool use, multimodal inputs, structured outputs, and agent-like workflows.
读寓言 →
Managed retrieval tool
文件搜索
File Search
File Search is a managed retrieval capability that indexes uploaded files and retrieves relevant content for model responses.
读寓言 →
Managed tool execution
代码解释器
Code Interpreter
Code Interpreter is a sandboxed tool that lets a model write and run code for analysis, computation, file manipulation, and visualization.
读寓言 →
Managed vector database
Pinecone
Pinecone
Pinecone is a managed vector database service for storing embeddings, filtering metadata, and serving similarity search for production AI applications.
读寓言 →
Marketing / Enterprise sales
标杆案例
Reference Case, Lighthouse Case, Benchmark Customer Case
标杆案例是可被公开或半公开复用的代表性客户成功样本，用来证明产品在某行业、某场景或某类客户中的可交付性和商业价值。
读寓言 →
MoE
辅助损失
Auxiliary Loss / Load Balancing Loss
An additional training loss that encourages uniform token distribution across experts, preventing routing collapse.
读寓言 →
MoE
专家路由
Expert Routing / Gating
The mechanism that decides which expert(s) each token should be processed by, typically a learned softmax gate over experts.
读寓言 →
MoE
Top-k 路由
Top-k Routing
A gating strategy where each token is sent to the top k experts rather than all experts, balancing utilization and quality.
读寓言 →
Model Risk Management
模型清单
Model Inventory
A centralized record of models in development, validation, production, retirement, or third-party use, including ownership and risk metadata.
读寓言 →
Model Risk Management
模型监控
Model Monitoring
Ongoing measurement of model behavior, performance, inputs, outputs, risks, and incidents after deployment.
读寓言 →
Model Risk Management
模型风险管理
Model Risk Management, MRM
The governance discipline for managing risks from model errors, misuse, drift, limitations, assumptions, and operational failures.
读寓言 →
Model Risk Management
模型验证
Model Validation
Independent or structured review that evaluates whether a model is fit for its intended purpose, performs as expected, and has known limitations.
读寓言 →
Model interface
函数调用
Function Calling
Function calling is a structured interface where a model emits machine-readable arguments for predefined functions so an application can safely execute external operations and return results.
读寓言 →
Model interface
JSON 模式
JSON Mode
A generation mode that biases or constrains output toward valid JSON.
读寓言 →
Model interface
结构化输出
Structured Output
Structured output constrains model responses into a specified schema such as JSON, enabling downstream systems to parse, validate, and execute results reliably.
读寓言 →
Model interface
结构化输出
Structured Outputs
Constraining model responses to a schema such as JSON for reliable downstream processing.
读寓言 →
Multi-agent architecture
监督智能体
Supervisor Agent
A supervisor agent monitors progress, assigns work, coordinates subagents, handles failures, and decides when to replan, escalate, or consolidate results.
读寓言 →
Multi-agent coordination
群聊编排
Group Chat Orchestration
Group chat orchestration coordinates multiple agents in a shared conversation, choosing speakers, routing messages, and deciding when collaboration is complete.
读寓言 →
Multi-agent framework
AutoGen
AutoGen
AutoGen is a framework for building applications where multiple conversational agents cooperate, call tools, execute code, and coordinate through message passing.
读寓言 →
Multi-agent organization unit
Agent 团队
Crew
A crew is a configured group of agents, tasks, tools, and process rules that work together toward a defined outcome.
读寓言 →
Multi-agent role
用户代理 Agent
User Proxy Agent
A user proxy agent represents the human user in a multi-agent system, often approving actions, providing input, or executing code on the user's behalf.
读寓言 →
Multi-agent workflow framework
CrewAI
CrewAI
CrewAI is a framework for structuring AI workers as role-based agents that perform tasks through sequential, hierarchical, or process-like collaboration.
读寓言 →
Multimodal
CLIP
Contrastive Language-Image Pre-training
A dual-encoder model that maps images and text into a shared embedding space using contrastive learning, serving as a foundation for many VLMs.
读寓言 →
Multimodal
交叉模态对齐
Cross-Modal Alignment
Training techniques (contrastive, regression, multimodal attention) to align representations across different modalities like text and images.
读寓言 →
Multimodal
多模态模型
Multimodal Model
A model trained on and capable of processing multiple modalities (text, image, audio, video), often capable of generating across modalities.
读寓言 →
Multimodal AI
视频理解
Video Understanding
模型理解视频中的画面、动作、语音、字幕和时间关系。
读寓言 →
Multimodality
对比学习
Contrastive Learning
A training approach that learns representations by pulling related examples closer and pushing unrelated examples apart in embedding space.
读寓言 →
Multimodality
跨模态嵌入
Cross-Modal Embedding
A shared vector representation space where different modalities, such as images and text, can be compared or retrieved by semantic similarity.
读寓言 →
Multimodality
多模态
Multimodality
The ability of a model or system to process and/or generate more than one data modality, such as text, image, audio, video, or actions.
读寓言 →
Multimodality
视觉语言模型
Vision-Language Model, VLM
A multimodal model that connects visual inputs with language understanding or generation.
读寓言 →
NLP / RAG
实体抽取
Entity Extraction
从文本中识别人名、组织、地点、产品、事件等实体。
读寓言 →
NLP / knowledge graph
关系抽取
Relation Extraction
从文本中识别实体之间的关系，用于结构化知识构建。
读寓言 →
NLP and RAG framework
Haystack
Haystack
Haystack is an open-source framework for building search, question answering, RAG, and LLM pipelines over documents and enterprise knowledge.
读寓言 →
Observability
工具调用延迟
Tool Call Latency
外部工具从请求到返回的耗时，是 Agent 用户体验和吞吐的重要指标。
读寓言 →
Observability artifact
精修版跨度
Span
A span is one timed operation within a trace, such as a retrieval call, model completion, tool invocation, parser step, or database query.
读寓言 →
Observability artifact
精修版跟踪记录
Trace
A trace records the end-to-end execution path of a request, including model calls, tool calls, retrieval steps, latency, errors, and intermediate outputs.
读寓言 →
Observability standard
OpenTelemetry
OpenTelemetry, OTel
OpenTelemetry is a vendor-neutral standard for collecting traces, metrics, and logs, increasingly used to instrument LLM and agent applications.
读寓言 →
Open-source AI observability and evaluation tool
Arize Phoenix
Arize Phoenix
Arize Phoenix is an open-source tool for tracing, evaluating, and analyzing LLM, RAG, and machine learning application behavior.
读寓言 →
Open-source LLM observability platform
Langfuse
Langfuse
Langfuse is an open-source observability platform for tracing LLM calls, prompts, generations, scores, costs, and user feedback.
读寓言 →
Open-source vector database
Milvus
Milvus
Milvus is an open-source vector database designed for scalable similarity search over large embedding collections.
读寓言 →
Open-source vector database
Qdrant
Qdrant
Qdrant is an open-source vector database and search engine focused on vector similarity search with payload filtering and production APIs.
读寓言 →
Open-source vector database
Weaviate
Weaviate
Weaviate is an open-source vector database that combines vector search, metadata filtering, hybrid search, and schema-based object storage.
读寓言 →
Operations
变更管理
Change Management
The controlled process for requesting, reviewing, approving, testing, communicating, and deploying changes to AI systems, prompts, tools, models, or policies.
读寓言 →
Operations
精修版升级路径
Escalation Path
A predefined route for sending exceptions, high-risk decisions, incidents, or ambiguous cases to the right human authority.
读寓言 →
Operations
精修版异常处理
Exception Handling
The process for detecting, routing, resolving, documenting, and learning from cases where an AI system cannot proceed safely or confidently.
读寓言 →
Operations
热加载
Hot Reload
在不中断服务或少中断服务的情况下更新配置、提示词、工具或模型版本。
读寓言 →
Operations
精修版人工旁路监督
Human-on-the-Loop, HOTL
An operating pattern where AI can act automatically but humans monitor, intervene, pause, or override when needed.
读寓言 →
Operations
精修版事后复盘
Postmortem
A structured review after an incident or major failure to identify root causes, impacts, corrective actions, and prevention measures.
读寓言 →
Operations
质量保障
Quality Assurance, QA
The process of evaluating AI outputs, workflows, and operational results against defined quality standards before or after delivery.
读寓言 →
Operations
精修版回滚
Rollback
当新版本异常时恢复到旧版本或旧配置。
读寓言 →
Operations
精修版回滚计划
Rollback Plan
A predefined plan to restore a previous safe state if a model, prompt, workflow, or system change causes unacceptable behavior.
读寓言 →
Operations
精修版根因分析
Root Cause Analysis, RCA
找出问题背后真正原因，而不只处理表面症状。
读寓言 →
Operations / finance
云成本运营
FinOps
FinOps 是跨工程、财务和业务团队管理云资源成本、预算、归因和优化的实践体系。
读寓言 →
Operations / infrastructure
自动扩缩容
Autoscaling
自动扩缩容根据负载、队列长度、CPU、GPU 或请求指标动态调整资源数量，以平衡成本和性能。
读寓言 →
Operations / platform
大语言模型运维
LLMOps
LLMOps 是面向大语言模型应用的运维体系，重点管理提示词、检索、评测、安全、成本、延迟和模型版本。
读寓言 →
Operations / platform
机器学习运维
MLOps
MLOps 是将机器学习模型从实验推进到稳定生产的工程体系，覆盖数据、训练、部署、监控和治理。
读寓言 →
Operations / reliability
死信队列
Dead-Letter Queue, DLQ
死信队列用于保存多次处理失败的消息或任务，便于后续排查、修复和重放。
读寓言 →
Operations / reliability
灾难恢复
Disaster Recovery, DR
灾难恢复是在重大故障、数据丢失或区域不可用时恢复服务和数据的计划、流程和技术能力。
读寓言 →
Operations / reliability
精修版可观测性
Observability
可观测性是通过日志、指标、追踪和事件来理解系统内部状态并定位问题的能力。
读寓言 →
Operations / reliability
恢复点目标
Recovery Point Objective, RPO
RPO 是灾难恢复中可接受的数据丢失时间窗口，决定备份和复制频率要求。
读寓言 →
Operations / reliability
恢复时间目标
Recovery Time Objective, RTO
RTO 是灾难或故障发生后业务必须恢复到可接受状态的最长时间。
读寓言 →
Operations / security
精修版事故响应
Incident Response
对系统故障、安全事件或严重质量问题进行检测、止血、恢复和复盘。
读寓言 →
Operations/Audit
抽样复核
Sampling Review
Reviewing a representative or risk-based sample of AI outputs or actions to estimate quality, compliance, and control effectiveness.
读寓言 →
Operations/Auditability
版本控制
Version Control
Tracking changes to prompts, models, datasets, tools, policies, code, and configuration so past states can be reviewed or restored.
读寓言 →
Operations/Governance
精修版审批门槛
Approval Gate
A defined point in a workflow where human approval or policy validation is required before the AI system proceeds.
读寓言 →
Operations/Governance
变更咨询委员会
Change Advisory Board, CAB
A cross-functional review group that assesses significant changes for risk, readiness, compliance, and operational impact.
读寓言 →
Operations/Risk
AI 事故
AI Incident
An event where an AI system causes or nearly causes harm, policy violation, security issue, compliance failure, or material operational disruption.
读寓言 →
Operations/Risk
业务连续性
Business Continuity
The ability to continue critical operations during disruptions affecting AI systems, vendors, infrastructure, data, or staffing.
读寓言 →
Operations/Risk
置信度阈值
Confidence Threshold
A predefined threshold used to decide whether an AI output can proceed automatically or must be reviewed, rejected, or escalated.
读寓言 →
Operations/Risk
精修版事实核验
Fact Verification
The process of checking AI-generated claims against trusted sources, evidence, or authoritative systems before use.
读寓言 →
Operations/Security
事故管理
Incident Management
The process for detecting, triaging, containing, investigating, communicating, and remediating AI, security, privacy, or operational incidents.
读寓言 →
Orchestration
有向无环图
Directed Acyclic Graph, DAG
A graph structure used to model step dependencies without cycles.
读寓言 →
Orchestration
意图识别
Intent Classification
判断用户请求所属类别，以决定后续流程、工具或 Agent。
读寓言 →
Orchestration
精修版语义路由
Semantic Routing
根据请求语义把任务分发到不同模型、工具、Agent 或工作流。
读寓言 →
Orchestration component
精修版路由器
Router
A router directs a request or intermediate task to the most suitable model, tool, workflow, skill, or agent based on intent, capability, cost, latency, and policy.
读寓言 →
PEFT
适配器
Adapter
Small bottleneck modules inserted between existing layers that have few trainable parameters; the original model weights remain frozen.
读寓言 →
PEFT
IA3
Infused Adapter by Inhibiting and Amplifying Inner Activations
An ultra-parameter-efficient method that scales key, value, and FFN activations by learned vectors, requiring fewer parameters than LoRA.
读寓言 →
PEFT
合并权重
Merge Weights / Merge Adapters
Fusing LoRA weights back into the base model for deployment, eliminating inference overhead from adapter computation.
读寓言 →
PEFT
前缀微调
Prefix Tuning
Prepending learnable continuous "prefix" vectors to the input or to each transformer layer while keeping model weights frozen.
读寓言 →
PEFT
提示微调
Prompt Tuning
Learning soft prompt embeddings prepended to input tokens rather than hand-crafting discrete prompts; extremely parameter-efficient.
读寓言 →
PEFT
量化低秩适配
Quantized LoRA, QLoRA
Combining 4-bit quantization of the base model with LoRA adapters to enable fine-tuning large models on consumer GPUs.
读寓言 →
PEFT
目标模块选择
Target Module Selection
The practice of choosing which layers (attention, FFN, all) to apply LoRA/adapters to, significantly affecting adaptation quality and efficiency.
读寓言 →
Platform / backend infrastructure
精修版工作流引擎
Workflow Engine
工作流引擎负责持久化、调度和恢复多步骤业务流程，保证任务在失败、重试和长时间运行下仍可完成。
读寓言 →
Platform / billing
配额
Quota
配额定义用户、租户或计划在一定周期内可使用的资源上限，例如 token、请求数、存储量或并发数。
读寓言 →
Platform / integration
连接器
Connector
连接器是把 AI 系统接入外部应用、数据库、文件、SaaS 或企业系统的适配组件。
读寓言 →
Platform / reliability
精修版持久执行
Durable Execution
持久执行通过记录工作流状态和事件，使长任务即使进程崩溃、重启或超时也能继续或恢复。
读寓言 →
Platform architecture
云原生
Cloud Native
云原生是一组利用容器、弹性伸缩、声明式基础设施和自动化运维构建可扩展系统的方法。
读寓言 →
Platform engineering
内部开发者平台
Internal Developer Platform, IDP
内部开发者平台为工程团队提供标准化的创建、部署、监控和运维应用的自助入口。
读寓言 →
Platform engineering / governance
黄金路径
Golden Path, Paved Road
黄金路径是平台团队为常见开发和部署场景提供的推荐流程、模板和默认配置，帮助团队快速且合规地交付。
读寓言 →
Policy
AI 政策
AI Policy
A formal organizational policy defining permitted, restricted, and prohibited AI uses, required approvals, data handling rules, and accountability expectations.
读寓言 →
Policy
可接受使用政策
Acceptable Use Policy, AUP
A policy that defines how employees, contractors, or systems may use AI tools and what behaviors or data inputs are prohibited.
读寓言 →
PostgreSQL vector extension
pgvector
pgvector
pgvector adds vector storage and similarity search to PostgreSQL so teams can combine embeddings with relational data and SQL workflows.
读寓言 →
Preprocessing
BPE
Byte-Pair Encoding
A subword tokenization algorithm that iteratively merges the most frequent character/byte pairs to build a vocabulary.
读寓言 →
Preprocessing
SentencePiece
SentencePiece
A language-independent tokenizer treating input as raw Unicode, supporting BPE and Unigram models; used by Llama, T5, and many others.
读寓言 →
Pretraining
语料库
Corpus
A collection of text, code, images, audio, or other data used to train or evaluate models.
读寓言 →
Pretraining
基座模型
Foundation Model
A large model trained on broad data that can be adapted to many downstream tasks.
读寓言 →
Pretraining
预训练
Pretraining
The initial large-scale training phase where a model learns broad patterns from massive datasets before task-specific adaptation.
读寓言 →
Pretraining
自监督学习
Self-Supervised Learning
A learning approach where supervisory signals are created from the data itself, such as predicting masked or next tokens.
读寓言 →
Privacy
匿名化
Anonymization
Irreversible transformation of data so individuals can no longer be identified by reasonable means.
读寓言 →
Privacy
数据最小化
Data Minimization
The principle of collecting, processing, retaining, and exposing only the data necessary for a defined purpose.
读寓言 →
Privacy
数据保护影响评估
Data Protection Impact Assessment, DPIA
A privacy assessment used to identify and mitigate risks when processing personal data, especially for high-risk or sensitive processing.
读寓言 →
Privacy
数据保留政策
Data Retention Policy
A policy defining how long data is stored, when it must be deleted or archived, and who approves exceptions.
读寓言 →
Privacy
差分隐私
Differential Privacy
通过加入数学噪声限制单个样本对输出的影响，以降低隐私泄露风险。
读寓言 →
Privacy
成员推断攻击
Membership Inference Attack
推断某条数据是否出现在模型训练集中，从而带来隐私风险。
读寓言 →
Privacy
模型反演攻击
Model Inversion Attack
通过模型输出反推出训练数据或敏感特征。
读寓言 →
Privacy
个人数据
Personal Data
Any information relating to an identified or identifiable natural person, as commonly defined in privacy regulations such as GDPR.
读寓言 →
Privacy
个人身份信息
Personally Identifiable Information, PII
Information that can identify, contact, locate, or distinguish a specific individual either directly or when combined with other data.
读寓言 →
Privacy
假名化
Pseudonymization
Processing personal data so it can no longer be attributed to a person without separately held additional information.
读寓言 →
Privacy
目的限制
Purpose Limitation
The principle that personal data should be collected for specified purposes and not reused in incompatible ways without proper basis.
读寓言 →
Privacy
安全聚合
Secure Aggregation
在联邦学习等场景中聚合参与方更新，同时隐藏单方明文数据。
读寓言 →
Privacy
敏感个人数据
Sensitive Personal Data
Personal data requiring heightened protection because misuse could create significant harm, discrimination, or legal exposure.
读寓言 →
Privacy/Compliance
跨境数据传输
Cross-Border Data Transfer
The movement or remote access of data across national or regional boundaries, often requiring legal safeguards and transfer mechanisms.
读寓言 →
Privacy/Compliance
数据本地化
Data Localization
A legal or policy requirement that certain data must remain within a specific country or jurisdiction.
读寓言 →
Privacy/Compliance
数据驻留
Data Residency
Requirements or commitments about the geographic location where data is stored, processed, replicated, or backed up.
读寓言 →
Procurement
供应商评估
Vendor Evaluation
The procurement and risk review process for assessing an AI vendor's capabilities, security, compliance, economics, support, and operational fit.
读寓言 →
Procurement/Architecture
供应商锁定
Vendor Lock-In
Dependency risk created when switching away from a vendor is costly due to proprietary APIs, data formats, workflows, contracts, or model-specific behavior.
读寓言 →
Procurement/Privacy
数据处理协议
Data Processing Agreement, DPA
A contract governing how a vendor or processor handles personal data on behalf of a customer or controller.
读寓言 →
Procurement/Privacy
数据使用权
Data Usage Rights
Contractual terms defining whether and how a vendor may use customer data for service delivery, analytics, product improvement, training, or model tuning.
读寓言 →
Procurement/Privacy
子处理方
Subprocessor
A third party engaged by a processor to process personal data as part of delivering a service.
读寓言 →
Procurement/Privacy
训练数据排除
Training Data Exclusion
A contractual or technical commitment that customer data will not be used to train or improve vendor models without permission.
读寓言 →
Procurement/Risk
退出计划
Exit Plan
A plan for safely terminating or replacing a vendor while preserving data access, service continuity, compliance, and operational control.
读寓言 →
Procurement/Risk
第三方风险管理
Third-Party Risk Management, TPRM
The lifecycle discipline for assessing, contracting, monitoring, and offboarding vendors and partners that introduce operational, security, privacy, or compliance risk.
读寓言 →
Procurement/Risk
供应商尽职调查
Vendor Due Diligence
The structured collection and review of evidence about a vendor's security posture, privacy practices, financial stability, compliance, and service reliability.
读寓言 →
Procurement/Security
安全问卷
Security Questionnaire
A structured questionnaire used by buyers to assess a vendor's security controls, compliance posture, data handling, and incident practices.
读寓言 →
Product / business strategy
产品市场匹配
Product-Market Fit, PMF
产品市场匹配是产品为明确市场群体解决强需求，并表现出留存、复购、推荐或收入增长信号的状态。
读寓言 →
Product / customer success
价值实现时间
Time to Value, TTV
价值实现时间是客户从开始使用到获得可感知业务价值所需的时间，越短通常越利于转化和留存。
读寓言 →
Product / platform
自助服务
Self-Service
自助服务让用户或内部团队无需人工介入即可完成注册、配置、部署、查询、升级或排障等操作。
读寓言 →
Product / research
用户画像
Persona, User Persona
用户画像是对典型使用者目标、任务、痛点、环境和行为模式的抽象描述，用于指导产品设计和沟通。
读寓言 →
Product / sales
概念验证
Proof of Concept, POC
概念验证是在有限范围内验证技术可行性、业务价值或采购信心的试点项目。
读寓言 →
Product engineering / release management
特性开关
Feature Flag, Feature Toggle
特性开关允许在不重新部署代码的情况下为不同用户、租户或环境开启或关闭功能。
读寓言 →
Product experience / customer success
入门引导
Onboarding
入门引导帮助新用户或客户完成账号配置、数据导入、首次任务和价值确认，以缩短上手时间。
读寓言 →
Product growth
激活
Activation
激活是新用户首次体验到产品核心价值的过程或关键行为。
读寓言 →
Product strategy
最小可行产品
Minimum Viable Product, MVP
MVP 是用最小产品范围验证核心用户需求、价值主张或商业假设的早期交付物。
读寓言 →
Product strategy / research
待完成任务
Jobs to Be Done, JTBD
JTBD 关注用户在特定情境下想取得的进展，而不是只按人口属性或功能偏好理解需求。
读寓言 →
Product strategy / workflow design
AI 原生工作流
AI-Native Workflow
AI 原生工作流不是把模型嵌入旧界面，而是围绕 AI 的理解、生成、调用工具和持续学习能力重新设计任务流程。
读寓言 →
Prompt optimization tool
提示优化器
Teleprompter
A DSPy teleprompter is an optimizer that searches for effective prompts, demonstrations, or program settings using examples and evaluation metrics.
读寓言 →
Prompt/function abstraction
语义函数
Semantic Function
A semantic function wraps a prompt or natural-language instruction as a callable unit that can be composed with native code functions.
读寓言 →
Quality assurance
精修版评测框架
Evaluation Harness
An evaluation harness runs repeatable tests against agent behavior, tool calls, outputs, latency, and failure cases to measure quality before and after changes.
读寓言 →
Quantization
双重量化
Double Quantization
Quantizing the quantization scaling constants themselves for an additional ~0.4 bits per parameter savings; introduced in QLoRA.
读寓言 →
Quantization
NF4
NormalFloat 4-bit
An information-theoretically optimal 4-bit data type for normally distributed weights, introduced with QLoRA.
读寓言 →
RAG
多模态 RAG
Multimodal RAG
同时检索文本、图片、表格、音频或视频证据来增强生成。
读寓言 →
RAG / search
重排序器
Reranker
重排序器会对初步检索结果重新评分排序，以提高进入模型上下文的信息相关性。
读寓言 →
RAG evaluation
精修版答案相关性
Answer Relevance
衡量模型回答是否真正回应用户问题，而不是只给出看似合理的文本。
读寓言 →
RAG evaluation
精修版上下文精确率
Context Precision
衡量提供给模型的上下文中相关信息的密度。
读寓言 →
RAG evaluation
精修版上下文召回率
Context Recall
衡量答案所需信息是否被包含在检索上下文中。
读寓言 →
RAG evaluation
精修版忠实性
Faithfulness
衡量生成答案是否受到给定上下文支持，常用于检测幻觉。
读寓言 →
RAG evaluation
检索精确率
Retrieval Precision
衡量检索结果中真正相关内容所占比例，避免把噪声塞进上下文。
读寓言 →
RAG evaluation
检索召回率
Retrieval Recall
衡量检索系统是否找回了应该被找回的相关证据，是 RAG 质量的基础指标。
读寓言 →
Reasoning
思维图
Graph of Thoughts
用图结构组织中间想法，使推理可合并、回溯和重用。
读寓言 →
Reasoning
自我一致性
Self-Consistency
对同一问题采样多条推理路径并投票选择答案，以提高复杂推理稳定性。
读寓言 →
Reasoning
思维树
Tree of Thoughts
将推理过程展开成多条可搜索路径，并通过评估选择更优分支。
读寓言 →
Reasoning / tool use
程序辅助语言模型
Program-Aided Language Model, PAL
让模型生成并执行程序来解决数学、逻辑或数据处理问题。
读寓言 →
Reasoning/action pattern
推理-行动范式
ReAct
ReAct is a prompting and agent pattern that interleaves reasoning traces with concrete actions, allowing the agent to decide what to do, call tools, observe results, and continue.
读寓言 →
Regulation
EU AI 法案
EU AI Act
The European Union regulation establishing risk-based obligations for AI systems, including prohibited practices, high-risk requirements, transparency duties, and GPAI obligations.
读寓言 →
Regulation
通用目的 AI
General-Purpose AI, GPAI
An AI model with broad capabilities that can be integrated into many downstream systems or tasks, subject to specific transparency and risk obligations under the EU AI Act.
读寓言 →
Regulation
高风险 AI 系统
High-Risk AI System
Under the EU AI Act, an AI system used in specified sensitive domains or safety contexts that must meet stricter risk, data, documentation, oversight, and monitoring requirements.
读寓言 →
Regulation/Compliance
合格评定
Conformity Assessment
A formal process to demonstrate that a regulated AI system meets applicable legal or technical requirements before market placement or use.
读寓言 →
Reliability / AI platform
回退
Fallback
回退是在主路径失败、超时或质量不足时自动切换到备用模型、服务、缓存或人工流程。
读寓言 →
Reliability / architecture
熔断器
Circuit Breaker
熔断器是在下游服务持续失败时临时阻止继续调用，以避免级联故障并给系统恢复时间。
读寓言 →
Reliability / operations
错误预算
Error Budget
错误预算是在 SLO 允许范围内可消耗的不可靠额度，用于平衡发布速度和系统稳定性。
读寓言 →
Reliability / operations
服务级别目标
Service Level Objective, SLO
SLO 是系统在可用性、延迟、错误率等方面承诺达到的可衡量目标。
读寓言 →
Reliability / product experience
降级
Graceful Degradation
降级是在依赖异常或资源紧张时保留核心功能、减少非关键能力或切换到低成本替代路径。
读寓言 →
Reliability pattern
模型回退
Model Fallback
Model fallback routes a request to an alternate model or provider when the primary option fails, times out, exceeds budget, or violates policy.
读寓言 →
Reliability technique
事实锚定
Grounding
Grounding ties an agent's outputs and decisions to supplied evidence, tool results, source documents, or verified system state rather than unsupported model guesses.
读寓言 →
Representation learning
嵌入
Embeddings
Embeddings are dense numerical representations of text, images, audio, code, or entities that place semantically similar items near each other in vector space.
读寓言 →
Reranking
交叉编码器
Cross-Encoder
将查询和候选文档一起输入模型进行精细相关性打分。
读寓言 →
Retrieval
BM25
BM25
基于词频和逆文档频率的经典稀疏检索算法。
读寓言 →
Retrieval
双编码器
Bi-Encoder
分别编码查询和文档以便高效向量相似度检索。
读寓言 →
Retrieval
假设文档嵌入
Hypothetical Document Embeddings, HyDE
先生成一个假设答案或文档，再用其嵌入进行检索。
读寓言 →
Retrieval
最大边际相关性
Maximal Marginal Relevance, MMR
在相关性和多样性之间取平衡，减少检索结果重复。
读寓言 →
Retrieval
多跳检索
Multi-Hop Retrieval
需要跨多个证据片段或实体关系逐步检索才能回答问题。
读寓言 →
Retrieval
查询改写
Query Rewriting
将用户原始问题改写成更适合检索或执行的查询。
读寓言 →
Retrieval
递归检索
Recursive Retrieval
根据初次检索结果继续展开子查询或相关节点，逐层补充上下文。
读寓言 →
Retrieval infrastructure
近似最近邻搜索
Approximate Nearest Neighbor Search (ANN Search)
ANN search finds vectors that are close enough to a query vector much faster than exact nearest-neighbor search, trading a small amount of recall for speed and scale.
读寓言 →
Retrieval infrastructure abstraction
文档存储
Document Store
A document store holds text, metadata, embeddings, and search indexes used by retrieval and question-answering pipelines.
读寓言 →
Retrieval infrastructure ecosystem
向量数据库生态
Vector Database Ecosystem
The vector database ecosystem includes specialized vector stores, search engines, libraries, and managed services for embedding-based retrieval.
读寓言 →
Retrieval method
密集检索
Dense Retrieval
Dense retrieval uses neural embeddings to compare query and document meaning through vector similarity, allowing retrieval even when wording differs.
读寓言 →
Retrieval method
混合搜索
Hybrid Search
Hybrid search combines lexical matching, such as BM25, with vector similarity search to capture both exact keyword intent and semantic meaning.
读寓言 →
Retrieval method
语义搜索
Semantic Search
Semantic search retrieves results based on meaning rather than exact keyword overlap, typically by embedding queries and documents into the same vector space.
读寓言 →
Retrieval method
稀疏检索
Sparse Retrieval
Sparse retrieval represents text with high-dimensional sparse features such as terms or learned sparse weights, making it strong for exact-match, rare-token, and keyword-sensitive queries.
读寓言 →
Retrieval preparation
分块
Chunking
Chunking splits source documents into retrieval-sized units so a RAG system can index and retrieve the most relevant pieces without overloading the context window.
读寓言 →
Retrieval preparation
父子分块
Parent-Child Chunking
用小块做匹配、返回更大的父块来兼顾命中精度和上下文完整性。
读寓言 →
Retrieval preparation
语义分块
Semantic Chunking
按语义边界而不是固定长度切分文档。
读寓言 →
Retrieval preparation
滑动窗口分块
Sliding Window Chunking
用重叠窗口切分长文档，降低跨块语义断裂风险。
读寓言 →
Retrieval quality
重排序
Reranking
Reranking applies a stronger relevance model after initial retrieval to reorder candidate results and improve the quality of evidence passed to the generation model.
读寓言 →
Retrieval/application abstraction
查询引擎
Query Engine
In LlamaIndex and similar frameworks, a query engine takes a user query, retrieves relevant data, and synthesizes an answer through an LLM.
读寓言 →
Risk Management
AI 风险管理
AI Risk Management
The systematic process of identifying, measuring, prioritizing, mitigating, monitoring, and reporting risks created by AI systems across their lifecycle.
读寓言 →
Risk Management
AI 风险管理框架
AI Risk Management Framework, AI RMF
A structured framework used to govern AI risks, commonly organized around functions such as govern, map, measure, and manage.
读寓言 →
Risk Management
幻觉风险
Hallucination Risk
The risk that a generative AI system produces false, unsupported, or fabricated information with convincing presentation.
读寓言 →
Risk Management
剩余风险
Residual Risk
The risk that remains after controls and mitigations have been applied and accepted or escalated by accountable owners.
读寓言 →
Risk Management
风险接受
Risk Acceptance
A formal decision by an authorized owner to proceed with a known residual risk within defined limits and monitoring requirements.
读寓言 →
Risk Management
风险偏好
Risk Appetite
The level and types of AI-related risk an organization is willing to accept in pursuit of business objectives.
读寓言 →
Risk Management
风险评估
Risk Assessment
The evaluation of an AI use case or system to determine potential harms, compliance obligations, likelihood, severity, and required controls.
读寓言 →
Risk Management
风险所有人
Risk Owner
The accountable person or function responsible for managing a specific risk within approved thresholds.
读寓言 →
Risk Management
风险登记册
Risk Register
A living inventory of identified risks, owners, likelihood, impact, mitigations, residual exposure, and review status.
读寓言 →
SaaS / platform architecture
多租户
Multi-Tenancy
多租户是让多个客户或组织共享同一套平台能力，同时在数据、权限、配置和计费上保持隔离。
读寓言 →
SaaS / security architecture
租户隔离
Tenant Isolation
租户隔离确保不同客户的数据、配置、权限和资源不会相互泄露或错误影响。
读寓言 →
Safety and Alignment
安全护栏
Guardrails
Rules, filters, classifiers, or system constraints used to reduce unsafe, invalid, or undesired model behavior.
读寓言 →
Sales AI / Revenue operations
销售智能体
Sales Agent, AI Sales Agent, Sales Copilot
销售智能体是围绕线索挖掘、客户研究、外呼或触达、话术生成、CRM 更新、商机推进和销售复盘等任务设计的 AI 执行系统。
读寓言 →
Security
AI 安全
AI Security
The discipline of protecting AI systems, models, data, prompts, tools, and agent actions from misuse, compromise, leakage, or adversarial manipulation.
读寓言 →
Security
基于属性的访问控制
Attribute-Based Access Control, ABAC
Access control based on attributes of users, resources, environment, and policies.
读寓言 →
Security
后门攻击
Backdoor Attack
在模型或数据中植入隐藏触发条件，使系统在特定输入下产生攻击者期望的输出。
读寓言 →
Security
上下文污染
Context Poisoning
恶意或低质量内容进入上下文后影响模型判断或行为。
读寓言 →
Security
数据外泄
Data Exfiltration
Unauthorized transfer of data out of a system, organization, or controlled boundary, including through agent tools or malicious prompts.
读寓言 →
Security
数据投毒
Data Poisoning
在训练、微调、索引或知识库中注入恶意数据以改变模型或系统行为。
读寓言 →
Security
越权代理
Excessive Agency
Risk caused by giving an AI agent too much autonomy, tool access, or permission scope.
读寓言 →
Security
模型投毒
Model Poisoning
攻击模型参数、权重或适配器，使其产生后门或错误行为。
读寓言 →
Security
OWASP LLM Top 10
OWASP Top 10 for LLM Applications
A risk taxonomy for common LLM application vulnerabilities.
读寓言 →
Security
权限边界
Permission Boundary
Explicit limits on what an agent or tool is allowed to read, write, call, or modify.
读寓言 →
Security
基于角色的访问控制
Role-Based Access Control, RBAC
Assigning permissions to roles rather than individual identities.
读寓言 →
Security
供应链风险
Supply Chain Risk
Risk introduced through dependencies, models, datasets, tools, plugins, or external services.
读寓言 →
Security
工具输出污染
Tool Output Poisoning
外部工具返回恶意指令或污染数据，诱导 Agent 偏离原任务。
读寓言 →
Security / agent runtime
沙箱
Sandbox
沙箱是隔离代码、文件、网络或工具执行的受控环境，用于限制 AI agent 或不可信任务的影响范围。
读寓言 →
Security / compliance
审计日志
Audit Log
审计日志记录谁在何时对哪些资源执行了什么操作，用于追责、合规、回放和故障分析。
读寓言 →
Security / governance
权限模型
Permission Model
权限模型定义用户、agent、工具和资源之间允许执行哪些操作，是防止越权和误操作的基础。
读寓言 →
Security / operations
最小权限原则
Principle of Least Privilege
最小权限原则要求系统组件或用户只拥有完成当前任务所需的最低权限，以降低事故和攻击影响。
读寓言 →
Security / operations
机密管理
Secrets Management
机密管理是安全存储、分发、轮换和审计 API key、密码、证书等敏感凭据的实践。
读寓言 →
Security/Compliance
信息安全管理体系
Information Security Management System, ISMS
The governance system of policies, controls, risk treatment, roles, and audits used to manage information security.
读寓言 →
Security/Governance
职责分离
Segregation of Duties, SoD
A control pattern that separates conflicting responsibilities so no one person or agent can complete a risky process without oversight.
读寓言 →
Security/Privacy
数据防泄漏
Data Loss Prevention, DLP
Tools and controls that detect, block, or monitor sensitive data movement across endpoints, networks, cloud services, and AI workflows.
读寓言 →
Serving
vLLM
vLLM
A high-throughput LLM serving engine using paged attention and continuous batching, the de facto standard for production LLM serving.
读寓言 →
Serving / UX
流式响应
Streaming Response
模型边生成边返回内容，降低用户感知等待时间。
读寓言 →
Serving metric
每秒 token 数
Tokens Per Second, TPS
衡量模型生成速度和推理吞吐的指标。
读寓言 →
Serving optimization
Chunked Prefill
Chunked Prefill
将长输入预填充分块调度，避免长请求阻塞短请求。
读寓言 →
Serving optimization
草稿模型
Draft Model
投机解码中用于快速产生候选 token 的较小模型。
读寓言 →
Serving optimization
KV Cache
Key-Value Cache
缓存 Transformer 注意力层历史 token 的键和值，减少重复计算但消耗显存。
读寓言 →
Software architecture / workflow
状态机
State Machine
状态机用明确的状态和状态转移描述系统流程，适合管理审批、任务执行和复杂 UI 或 agent 生命周期。
读寓言 →
Speech AI
语音识别
Automatic Speech Recognition, ASR
将语音转换成文本。
读寓言 →
Speech AI
语音合成
Text-to-Speech, TTS
将文本转换成可听语音。
读寓言 →
System architecture
多智能体编排
Multi-Agent Orchestration
Multi-agent orchestration distributes work across specialized agents and coordinates their roles, dependencies, messages, and outputs to solve tasks that benefit from parallelism or separation of concerns.
读寓言 →
System architecture
编排
Orchestration
Orchestration coordinates models, tools, workflows, agents, human approvals, state, and error handling so a complete task can move through multiple components reliably.
读寓言 →
System design
上下文工程
Context Engineering
Context engineering is the discipline of selecting, organizing, compressing, and updating the information supplied to a model so it can act correctly within token, latency, and reliability constraints.
读寓言 →
Tokenization
字节对编码
Byte Pair Encoding, BPE
A common subword tokenization method that builds tokens by repeatedly merging frequent character or byte pairs.
读寓言 →
Tokenization
词元
Token
A unit of text or data representation used by a model, often a word piece, character sequence, or symbol.
读寓言 →
Tokenization
分词
Tokenization
The process of converting raw text into smaller units called tokens that a model can process.
读寓言 →
Tokenization
分词器
Tokenizer
The software component that maps text to token IDs and often maps generated token IDs back to text.
读寓言 →
Tokenization
词表
Vocabulary
The fixed set of tokens a tokenizer can map to model-readable IDs.
读寓言 →
Tool
浏览器工具
Browser Tool
让 Agent 打开网页、读取页面、点击或执行网页任务的工具。
读寓言 →
Tool interface
插件
Plugin
A packaged external capability that an AI system can discover and invoke.
读寓言 →
Training
激活检查点
Activation Checkpointing / Gradient Checkpointing
Trading compute for memory by recomputing intermediate activations during backpropagation instead of storing them.
读寓言 →
Training
bf16
Brain Floating Point 16
A 16-bit floating-point format with the same exponent range as fp32 but reduced mantissa, offering training stability without gradient scaling.
读寓言 →
Training
灾难性遗忘
Catastrophic Forgetting
The tendency of neural networks to lose previously learned capabilities when fine-tuned on new data distributions.
读寓言 →
Training
思维链训练
Chain-of-Thought Training
Fine-tuning on step-by-step reasoning traces to improve multi-step reasoning, math, and planning capabilities.
读寓言 →
Training
Chinchilla 方案
Chinchilla Optimal
The finding that for a given compute budget, it is optimal to scale model size and training tokens roughly equally, rather than training very large models on less data.
读寓言 →
Training
持续预训练
Continual Pre-training / Continued Training
Extending pre-training on new domain data or more recent data to update a model's knowledge without starting from scratch.
读寓言 →
Training
课程学习
Curriculum Learning
Training on increasingly difficult examples, analogous to how humans progress from simple to complex material.
读寓言 →
Training
数据混合
Data Mixing
The strategy of blending multiple data sources with specific weights during pre-training or fine-tuning to balance capabilities.
读寓言 →
Training
去噪目标
Denoising Objective
Training objectives where the model learns to reconstruct clean data from corrupted or noisy versions; used in BERT, diffusion, and some autoencoders.
读寓言 →
Training
GRPO
Group Relative Policy Optimization
A variant of PPO that uses group-relative rewards—comparing outputs within the same prompt batch—eliminating the need for a value function model.
读寓言 →
Training
掩码语言建模
Masked Language Modeling, MLM
A pre-training objective where random input tokens are masked and the model must predict them from bidirectional context.
读寓言 →
Training
混合精度训练
Mixed Precision Training
Training with lower-precision formats (fp16, bf16) for most operations while keeping master weights in fp32 to save memory and increase throughput.
读寓言 →
Training
多任务学习
Multi-Task Learning
Training a single model on multiple tasks simultaneously so shared representations benefit all tasks.
读寓言 →
Training
下一个词预测
Next-Token Prediction
The standard autoregressive pre-training objective: predict each subsequent token in a sequence given all previous tokens.
读寓言 →
Training
预训练
Pre-training
The initial large-scale training phase where a model learns general capabilities from massive corpora, typically via self-supervised objectives like next-token prediction.
读寓言 →
Training
强化微调
Reinforcement Fine-Tuning, RFT
Combining SFT with reinforcement learning (RLVR/GRPO) to develop reasoning capabilities beyond what demonstration data alone provides.
读寓言 →
Training
可验证奖励强化学习
Reinforcement Learning with Verifiable Rewards, RLVR
Using reinforcement learning with programmatically verifiable rewards (e.g., math correctness, code execution) rather than human preferences.
读寓言 →
Training
缩放定律
Scaling Laws
Empirical relationships showing that model loss predictably decreases with compute, data, and parameter count following power-law behavior.
读寓言 →
Training / privacy
联邦学习
Federated Learning
多方在不集中原始数据的情况下协同训练模型。
读寓言 →
Training and Inference
知识蒸馏
Knowledge Distillation
A method where a smaller student model learns to imitate a larger teacher model's behavior.
读寓言 →
Training data
数据增强
Data Augmentation
通过变换或生成样本扩充训练数据，提高泛化能力。
读寓言 →
Transformers
注意力机制
Attention Mechanism
A mechanism that lets a model dynamically weigh which input elements are most relevant when producing each output representation.
读寓言 →
Transformers
解码器
Decoder
The part of a sequence model that generates output tokens, often one token at a time conditioned on prior context.
读寓言 →
Transformers
编码器
Encoder
The part of a sequence model that converts input tokens into contextual representations.
读寓言 →
Transformers
多头注意力
Multi-Head Attention
A transformer component that runs several attention operations in parallel so the model can capture different relationship patterns at once.
读寓言 →
Transformers
位置编码
Positional Encoding
Information added to token representations so a transformer can account for token order.
读寓言 →
Transformers
自注意力
Self-Attention
An attention mechanism where tokens in the same sequence attend to one another to build contextual representations.
读寓言 →
Transformers
Transformer
Transformer
A neural network architecture based on attention mechanisms that became the dominant foundation for modern language and multimodal models.
读寓言 →
Trust and attribution
引文
Citations
Citations attach source references to generated claims so users can inspect the underlying evidence, verify factual statements, and trace answer provenance.
读寓言 →
User-facing product / Productivity
AI助手
AI Assistant, Copilot
AI助手是面向个人或团队的对话式、协作式 AI 工具，通常用于问答、写作、检索、分析、操作建议或辅助完成某类工作。
读寓言 →
Vector search library
FAISS
Facebook AI Similarity Search, FAISS
FAISS is a library for efficient similarity search and clustering of dense vectors, commonly used as a local vector index backend.
读寓言 →
Workflow automation / Orchestration
AI工作流
AI Workflow, Agentic Workflow
AI工作流是把模型调用、规则判断、工具执行、人工审批和业务系统集成编排成可重复运行的流程，用于把 AI 能力落到稳定产出上。
读寓言 →
交互机制
智能体协商
Agent Negotiation
Agent 之间通过反复提议与反提议就资源分配、任务分工或策略选择达成协议的交互过程。
读寓言 →
协作机制
多智能体反思
Multi-Agent Reflection
Agent 在完成阶段性工作后互相审查、提出批评或修改建议，通过集体反思迭代提升输出质量。
读寓言 →
协作机制
任务分解
Task Decomposition
将复杂任务递归拆分为可被单个 Agent 或子团队独立执行的子任务。
读寓言 →
协作模式
智能体辩论
Agent Debate
多个 Agent 对同一命题持不同立场进行辩论，通过对抗性论证提升推理质量或达成更优决策。
读寓言 →
协作模式
智能体群
Agent Swarm
大量轻量 Agent 通过局部规则和涌现行为完成全局任务的去中心化协作模式，灵感来自蚁群/蜂群。
读寓言 →
协作模式
多智能体共识
Multi-Agent Consensus
多个 Agent 通过投票、加权或迭代协商对决策结果达成一致的协议机制。
读寓言 →
协作模式
角色扮演多智能体
Role-Playing Multi-Agent
每个 Agent 被分配不同角色（如 CEO、工程师、分析师），通过角色化对话协作推进任务。
读寓言 →
协作现象
涌现行为
Emergent Behavior
由于 Agent 间简单的局部规则相互作用，在系统层面自发产生的复杂、非预先编程的集体行为。
读寓言 →
可观测性
Agent 决策追踪
Agent Decision Tracing
记录 Agent 在每一步推理中选择做什么动作、为什么选这个动作及其备选方案的过程。
读寓言 →
可观测性
合规日志
Compliance Logging
按法规要求保留所有 LLM 交互的完整记录，包括用户输入、模型输出、过滤决策和人工干预痕迹。
读寓言 →
可观测性
用户反馈闭环
Feedback Loop
收集终端用户的赞/踩、纠错或评分信号，回传至模型迭代和评估管道的持续改进机制。
读寓言 →
可观测性
幻觉检测
Hallucination Detection
自动检测模型输出中与给定上下文或既知事实不一致的虚构内容。
读寓言 →
可观测性
LLM 可观测性
LLM Observability
通过采集链路日志、指标和调用链，全面洞察 LLM 应用在生产环境中的行为、性能和质量。
读寓言 →
可观测性
延迟监控
Latency Monitoring
持续测量 LLM 推理首 Token 时间 (TTFT) 和总生成时间，识别性能瓶颈和体验劣化。
读寓言 →
可观测性
生产漂移检测
Production Drift Detection
检测模型输入分布、输出质量或行为模式从训练/基线状态发生显著变化的自动化机制。
读寓言 →
可观测性
提示版本管理
Prompt Versioning
对每次部署的提示模板进行版本标记，关联线上效果，支持回滚与 A/B 对比。
读寓言 →
可观测性
Token 用量监控
Token Usage Monitoring
实时跟踪每次 API 调用的 Token 消耗量，用于成本控制、预算预警和效率优化。
读寓言 →
可观测性
调用链追踪
Tracing
记录一次完整请求在各 Agent、API、工具调用之间的调用路径、耗时和中间结果。
读寓言 →
基准
BIG-bench (超越模仿游戏基准)
Beyond the Imitation Game Benchmark
Google 牵头的超大规模协作基准，包含 200+ 任务，探测 LLM 在推理、伦理、幽默等维度的极限。
读寓言 →
基准
HumanEval (代码生成评估)
HumanEval
OpenAI 提出的代码生成基准，包含 164 个手写编程题，用 pass@k 衡量模型生成正确代码的能力。
读寓言 →
基准
MMLU (大规模多任务语言理解)
Massive Multitask Language Understanding
涵盖 57 个学科的多选题测试集，是衡量 LLM 知识广度与推理能力的标杆基准。
读寓言 →
基准
SWE-bench
SWE-bench
用真实 GitHub Issue 和 PR 测试模型自动修复软件缺陷能力的工程级基准。
读寓言 →
基准框架
HELM (整体语言模型评估)
Holistic Evaluation of Language Models
Stanford 提出的多维评估框架，从准确性、校准、鲁棒性、公平性、效率等多个轴对模型进行全景评分。
读寓言 →
安全方法
模型安全评估
Model Safety Evaluation
用标准化测试集（如 HarmBench、SafetyBench）系统评估模型在各类安全场景下的拒答率和有害输出率。
读寓言 →
安全方法
红队测试
Red Teaming
组织专门团队系统性地攻击模型以发现安全漏洞、有害输出和未被预见的风险边界。
读寓言 →
安全机制
有害内容过滤
Harmful Content Filtering
在输入和输出两端部署分类器或规则引擎，拦截暴力、色情、仇恨言论等违规内容。
读寓言 →
对齐技术
基于人类反馈的强化学习
RLHF (Reinforcement Learning from Human Feedback)
用人类偏好数据训练奖励模型，再通过强化学习微调 LLM 使其输出与人类价值观对齐。
读寓言 →
对齐技术
安全微调
Safety Fine-Tuning
在预训练或指令微调后，用安全相关数据对模型进行额外训练以增强拒绝危险请求的能力。
读寓言 →
工具
评估套件
Eval Suite / Eval Harness
统一管理多个基准、执行大批量自动评估并汇总报告的工具框架。
读寓言 →
指标
任务成功率
Task Success Rate
Agent 系统在端到端执行中实际达成任务目标的占比，是衡量 Agent 实用性的核心指标。
读寓言 →
攻击类型
对抗性示例
Adversarial Examples
对输入进行人类不可察觉的微小扰动，使模型产生全错误的预测或分类。
读寓言 →
攻击类型
拒绝服务攻击
Denial-of-Service (DoS) on LLM
通过构造极端消耗算力的输入（超长上下文、递归生成请求）耗尽模型服务资源。
读寓言 →
攻击类型
间接提示注入
Indirect Prompt Injection
恶意指令不直接发给 LLM，而是隐藏在网页、邮件或文档等外部数据源中，当模型读取时被触发。
读寓言 →
攻击类型
越狱
Jailbreak
通过精心构造的提示绕过模型的安全对齐机制，诱使其产生产生原本拒绝的有害或违规内容。
读寓言 →
攻击类型
多模态越狱
Multi-modal Jailbreak
利用图像、音频等多模态输入绕过大语言模型的安全过滤，实现纯文本输入无法实现的越狱。
读寓言 →
攻击类型
提示注入
Prompt Injection
攻击者在输入中嵌入恶意指令，覆盖或劫持系统提示的原始意图，使模型执行非预期的操作。
读寓言 →
攻击类型
越狱
Prompt Leaking
攻击者通过诱导性提问使模型泄露其系统提示、内部规则或私有上下文的攻击手法。
读寓言 →
架构
智能体编排
Agent Orchestration
定义多个 Agent 的调用顺序、数据流向、错误处理与状态传递的调度层。
读寓言 →
架构
黑板架构
Blackboard Architecture
多个专家 Agent 共享一个公共数据结构（黑板），各自在合适的时机读取和写入，协同解决复杂问题。
读寓言 →
架构
去中心化多智能体
Decentralized Multi-Agent
无中心调度节点，各 Agent 通过对等通信与局部决策协同完成任务。
读寓言 →
架构
层级多智能体
Hierarchical Multi-Agent
Agent 按树状或金字塔结构组织，上级 Agent 分配任务、汇总结果，下级 Agent 负责执行。
读寓言 →
架构
混合多智能体
Hybrid Multi-Agent
结合层级式调度与去中心化协作，不同子网采用不同组织方式以适应任务特性。
读寓言 →
架构
多智能体系统
Multi-Agent System (MAS)
由多个自主 Agent 组成、通过通信与协调共同完成任务的分布式系统。
读寓言 →
治理
AI 治理
AI Governance
企业或组织对 AI 系统开发、部署和运营建立的政策、流程、角色和问责框架的总称。
读寓言 →
治理
访问控制
Access Control
按角色、权限级别限制谁可以调用特定模型、使用敏感功能或访问特定知识库的机制。
读寓言 →
治理
Agent 身份与认证
Agent Identity & Authentication
赋予每个 Agent 唯一身份、为其签发凭证并验明其调用来源，实现 Agent 间和对外服务的可信通信。
读寓言 →
治理
审计日志
Audit Logs
不可篡改地记录每个 AI 决策的完整链——谁请求、模型做了什么、为什么做、是否经过人工审批。
读寓言 →
治理
自主性级别
Autonomy Levels
根据人类干预频率和决策权限范围对 AI 系统从完全手动到完全自主的分级，常见 0-5 或 L0-L5 分级。
读寓言 →
治理
数据隐私与 PII 遮蔽
Data Privacy & PII Masking
在输入 LLM 前自动识别并脱敏个人身份信息，在返回用户前按需还原的隐私保护机制。
读寓言 →
治理
人类在回路内
Human-in-the-Loop (Deep HITL)
人类不仅是最终审批者，而是作为协作参与者与 AI 系统在每一步紧密交互、共同决策的深度协作模式。
读寓言 →
治理
人类在回路中
Human-in-the-Loop (HITL)
在 AI 自动化流程的关键节点嵌入人工审核、干预或批准的机制，确保最终决策权在人类手中。
读寓言 →
治理
人类在回路外
Human-on-the-Loop (HOTL)
AI 系统自主执行但人类保持监督态势，仅在系统发出告警或异常时才介入的更宽松的控制模式。
读寓言 →
治理
模型审批流程
Model Approval Workflow
新模型或新能力上线前必须经过安全评估、风险审核和多级批准的正式流程。
读寓言 →
治理
模型风险评估
Model Risk Assessment
上线前对模型的能力边界、潜在有害使用场景、偏见放大风险和滥用可能性进行结构化评估。
读寓言 →
治理
速率限制
Rate Limiting
按用户或 API 密钥限制单位时间内的请求数量，防止滥用、控制成本和保障服务可用性。
读寓言 →
治理
负责任 AI
Responsible AI
确保 AI 系统在公平性、透明度、可解释性、隐私保护和安全性方面满足伦理和社会期望的实践体系。
读寓言 →
治理
后悔权与撤销机制
Right to Appeal & Rollback
用户对 AI 自动化决策有权请求人工复审或撤销，系统有机制回退已执行的自动化操作。
读寓言 →
治理
安全事故响应
Security Incident Response
当模型出现越狱、数据泄露或严重有害输出时，团队启动的应急止损、根因分析和修复流程。
读寓言 →
治理
使用策略
Usage Policy
明确定义可接受与禁止的模型使用场景、输入输出边界和执行规则的正式文档。
读寓言 →
评估
LLM 评估
LLM Evaluation (Eval)
通过自动化测试、人工评分或对抗性探针系统化衡量模型在特定任务上的能力与局限性。
读寓言 →
评估方法
对抗性评估
Adversarial Evaluation
故意构造具有迷惑性、边界性或恶意的输入来测试模型鲁棒性和安全边界的评估方法。
读寓言 →
评估方法
对齐评估
Alignment Evaluation
测量模型行为是否与人类意图、价值观和安全要求一致，包括有用性、诚实性和无害性三维度。
读寓言 →
评估方法
能力评估
Capability Evaluation
测量模型"能做什么"——推理、编码、工具调用、长上下文等正向能力的系统性评测。
读寓言 →
评估方法
人工评估
Human Evaluation
用一个（通常更强的）LLM 自动评判另一个 LLM 的输出质量，以替代部分人工评估。
读寓言 →
评估方法
逐轮评估
Per-Turn Evaluation
针对多轮对话场景，对每一轮对话的回复质量单独打分，而非仅做整段会话评分。
读寓言 →
评估问题
能力污染与数据泄露
Contamination & Data Leakage
训练数据中混入了测试集的题目或类似内容，导致基准分数虚高、不再反映真实泛化能力。
读寓言 →
评估问题
提示敏感度
Prompt Sensitivity
模型对同一个问题的不同措辞、格式或示例排列表现出显著差异的性能波动。
读寓言 →
调度
Agent 路由
Agent Routing
根据请求内容、Agent 能力描述和负载情况，将用户查询或子任务动态分配给最适合的 Agent。
读寓言 →
通信
Agent 通信协议
Agent Communication Protocol
定义 Agent 间消息格式、语义和交互规则的标准化语言或接口规范。
读寓言 →