Claude & AI Tech Glossary

Additional Categories

Deep Learning Fundamentals

Term	Full Form / Expansion	What It Means
Neural Network	—	A machine learning model composed of interconnected layers of artificial neurons that learn patterns from data.
Artificial Neuron	—	The basic computational unit of a neural network that applies weights, bias, and an activation function.
Deep Learning	—	A subset of machine learning that uses deep neural networks with many hidden layers to solve complex tasks.
Epoch	—	One complete pass through the entire training dataset during model training.
Batch	—	A subset of training samples processed together before updating model weights.
Batch Size	—	The number of training samples processed in a single forward and backward pass.
Mini Batch	—	A smaller subset of data used during gradient descent to improve training efficiency.
Gradient Descent	—	An optimization algorithm that minimizes the loss function by updating model parameters.
Stochastic Gradient Descent (SGD)	—	A variant of gradient descent that updates model weights using one or a few samples at a time.
Learning Rate	—	Controls how much model weights change during each optimization step.
Optimizer	—	Algorithm responsible for updating neural network weights during training (e.g., Adam, SGD, RMSProp).
Loss Function	—	Mathematical function that measures how incorrect the model's predictions are.
Cost Function	—	Aggregated loss across an entire dataset used to optimize model performance.
Backpropagation	—	Algorithm used to calculate gradients for updating neural network weights.
Activation Function	—	Function applied to neurons to introduce non-linearity (ReLU, GELU, Sigmoid, Tanh).
Forward Pass	—	The process of passing input data through a neural network to produce predictions.
Backward Pass	—	The process of propagating gradients backward through the network to update weights.
Overfitting	—	When a model memorizes training data instead of learning general patterns.
Underfitting	—	When a model is too simple to capture meaningful patterns in the data.
Regularization	—	Techniques used to reduce overfitting and improve model generalization.
Dropout	—	Regularization method that randomly disables neurons during training.
Normalization	—	Process of scaling data or activations to improve training stability.
Fine-Tuning	—	Continuing training of a pretrained model on a specialized dataset.
Pretraining	—	Initial large-scale training of a model on broad datasets before task-specific tuning.
Transfer Learning	—	Reusing knowledge from a pretrained model for another related task.

---

Inference & Model Serving

Term	Full Form / Expansion	What It Means
Inference Engine	—	Software responsible for efficiently executing trained AI models.
vLLM	Virtual Large Language Model	High-performance inference engine optimized for serving LLMs efficiently.
Ollama	—	Local LLM runtime for running and managing open-source models on personal hardware.
TensorRT-LLM	TensorRT Large Language Model	NVIDIA framework for optimizing LLM inference on GPUs.
MLX	Machine Learning eXplore	Apple's machine learning framework optimized for Apple Silicon.
GGUF	GPT-Generated Unified Format	Efficient file format used to store quantized language models.
GPTQ	GPT Quantization	Quantization technique that compresses LLM weights while preserving accuracy.
AWQ	Activation-aware Weight Quantization	Quantization method optimized for preserving model quality after compression.
Continuous Batching	—	Serving technique where requests are dynamically batched for better GPU utilization.
KV Cache Eviction	—	Removing old key-value cache entries to free memory during inference.
Model Sharding	—	Splitting a large model across multiple devices or servers.
Tensor Parallelism	—	Running different parts of a model simultaneously across multiple GPUs.
Pipeline Parallelism	—	Dividing model layers across multiple devices for sequential execution.
Data Parallelism	—	Replicating models across devices while splitting input data among them.
Model Serving	—	Deploying trained models so they can receive requests and return predictions.
Cold Start	—	Delay experienced when a model loads into memory for the first time.

---

RAG & Retrieval Systems

Term	Full Form / Expansion	What It Means
Dense Retrieval	—	Retrieval using embedding similarity rather than keyword matching.
Sparse Retrieval	—	Retrieval using traditional keyword-based methods like BM25.
BM25	Best Matching 25	A ranking algorithm widely used in traditional search engines.
Parent Document Retrieval	—	Retrieving small chunks while returning their larger parent documents for context.
Contextual Compression	—	Compressing retrieved documents to only include information relevant to the query.
Citation Generation	—	Producing references showing where retrieved information originated.
Retrieval Pipeline	—	Complete workflow from embedding generation to retrieval and reranking.
Query Expansion	—	Enhancing search queries with related terms to improve retrieval quality.
Retriever	—	Component responsible for finding relevant documents from a knowledge base.
Generator	—	LLM responsible for producing the final response using retrieved context.

---

Agent Frameworks

Term	Full Form / Expansion	What It Means
LangGraph	—	Graph-based framework for building stateful AI agent workflows.
CrewAI	—	Framework for orchestrating collaborative role-based AI agents.
AutoGen	—	Microsoft's framework for multi-agent conversations and automation.
Semantic Kernel	—	Microsoft SDK for integrating AI capabilities into applications.
Haystack	—	Open-source framework for search, RAG, and question-answering systems.
LlamaIndex	—	Framework for connecting LLMs to structured and unstructured data sources.
OpenAI Agents SDK	—	SDK for building AI agents using OpenAI models and tools.
Agent Workflow	—	Sequence of steps followed by one or more AI agents to complete a task.
Planner Agent	—	Agent responsible for task decomposition and execution planning.
Executor Agent	—	Agent responsible for performing planned tasks using tools and APIs.
Critic Agent	—	Agent responsible for evaluating and improving generated outputs.
Reflection	—	Process where an agent reviews and improves its own previous outputs.

---

Evaluation & LLMOps

Term	Full Form / Expansion	What It Means
LangSmith	—	Platform for tracing, debugging, evaluating, and monitoring LLM applications.
Arize Phoenix	—	Open-source AI observability platform for evaluating LLM performance.
Ragas	Retrieval-Augmented Generation Assessment	Framework for evaluating RAG system quality.
DeepEval	—	Open-source framework for evaluating LLM applications using automated metrics.
TruLens	—	Framework for evaluating and monitoring LLM applications.
Hallucination Rate	—	Percentage of generated responses containing fabricated information.
Faithfulness	—	Degree to which an answer is supported by retrieved evidence.
Relevance Score	—	Measure of how well retrieved information matches the user's query.
Prompt Versioning	—	Managing changes and versions of prompts over time.
Prompt Registry	—	Central repository for storing and managing production prompts.
Model Registry	—	Repository for storing and versioning trained AI models.
Experiment Tracking	—	Recording training runs, metrics, and configurations for reproducibility.

---

AI Security

Term	Full Form / Expansion	What It Means
Prompt Exfiltration	—	Attempt to extract hidden prompts, secrets, or confidential instructions.
Tool Hijacking	—	Manipulating an AI agent into using tools in unintended ways.
Indirect Prompt Injection	—	Malicious instructions hidden in retrieved documents or external websites.
Supply Chain Attack	—	Compromising AI systems through dependencies, models, or external components.
Model Poisoning	—	Introducing malicious data during model training or fine-tuning.
Data Leakage	—	Accidental exposure of confidential information by an AI system.
Secret Scanning	—	Detecting exposed credentials, API keys, and sensitive information.
AI Red Teaming	—	Security testing focused on discovering vulnerabilities in AI systems.
Jailbreak Detection	—	Identifying attempts to bypass model safety mechanisms.
Content Filtering	—	Screening AI inputs and outputs to enforce safety and policy requirements.
Policy Engine	—	Component that enforces organizational AI usage rules.
AI Risk Assessment	—	Evaluating potential security, privacy, and compliance risks of AI systems.

```