LangChain4j Glossary

Table of Contents

• Core Concepts
• LangChain4j Components
• AI/ML Concepts
• Guardrails
• Prompt Engineering
• RAG (Retrieval-Augmented Generation)
• Agents and Tools
• Agentic Module
• Model Context Protocol (MCP)
• Azure Services
• Testing and Development

Quick reference for terms and concepts used throughout the course.

Core Concepts

AI Agent - System that uses AI to reason and act autonomously. Module 04

Chain - Sequence of operations where output feeds into the next step.

Chunking - Breaking documents into smaller pieces. Typical: 300-500 tokens with overlap. Module 03

Context Window - Maximum tokens a model can process. GPT-5.2: 400K tokens (up to 272K input, 128K output).

Embeddings - Numerical vectors representing text meaning. Module 03

Function Calling - Model generates structured requests to call external functions. Module 04

Hallucination - When models generate incorrect but plausible information.

Prompt - Text input to a language model. Module 02

Semantic Search - Search by meaning using embeddings, not keywords. Module 03

Stateful vs Stateless - Stateless: no memory. Stateful: maintains conversation history. Module 01

Tokens - Basic text units models process. Affects costs and limits. Module 01

Tool Chaining - Sequential tool execution where output informs next call. Module 04

LangChain4j Components

AiServices - Creates type-safe AI service interfaces.

OpenAiOfficialChatModel - Unified client for OpenAI and Azure OpenAI models.

OpenAiOfficialEmbeddingModel - Creates embeddings using OpenAI Official client (supports both OpenAI and Azure OpenAI).

ChatModel - Core interface for language models.

ChatMemory - Maintains conversation history.

ContentRetriever - Finds relevant document chunks for RAG.

DocumentSplitter - Breaks documents into chunks.

EmbeddingModel - Converts text into numerical vectors.

EmbeddingStore - Stores and retrieves embeddings.

MessageWindowChatMemory - Maintains sliding window of recent messages.

PromptTemplate - Creates reusable prompts with {{variable}} placeholders.

TextSegment - Text chunk with metadata. Used in RAG.

ToolExecutionRequest - Represents tool execution request.

UserMessage / AiMessage / SystemMessage - Conversation message types.

AI/ML Concepts

Few-Shot Learning - Providing examples in prompts. Module 02

Large Language Model (LLM) - AI models trained on vast text data.

Reasoning Effort - GPT-5.2 parameter controlling thinking depth. Module 02

Temperature - Controls output randomness. Low=deterministic, high=creative.

Vector Database - Specialized database for embeddings. Module 03

Zero-Shot Learning - Performing tasks without examples. Module 02

Guardrails

Defense in Depth - Multi-layer security approach combining application-level guardrails with provider safety filters.

Hard Block - Provider throws HTTP 400 error for severe content violations.

InputGuardrail - LangChain4j interface for validating user input before it reaches the LLM. Saves cost and latency by blocking harmful prompts early.

InputGuardrailResult - Return type for guardrail validation: success() or fatal("reason").

OutputGuardrail - Interface for validating AI responses before returning to users.

Provider Safety Filters - Built-in content filters from AI providers (e.g., Azure OpenAI) that catch violations at the API level.

Soft Refusal - Model politely declines to answer without throwing an error.

Prompt Engineering - Module 02

Chain-of-Thought - Step-by-step reasoning for better accuracy.

Constrained Output - Enforcing specific format or structure.

High Eagerness - GPT-5.2 pattern for thorough reasoning.

Low Eagerness - GPT-5.2 pattern for quick answers.

Multi-Turn Conversation - Maintaining context across exchanges.

Role-Based Prompting - Setting model persona via system messages.

Self-Reflection - Model evaluates and improves its output.

Structured Analysis - Fixed evaluation framework.

Task Execution Pattern - Plan → Execute → Summarize.

RAG (Retrieval-Augmented Generation) - Module 03

Document Processing Pipeline - Load → chunk → embed → store.

In-Memory Embedding Store - Non-persistent storage for testing.

RAG - Combines retrieval with generation to ground responses.

Similarity Score - Measure (0-1) of semantic similarity.

Source Reference - Metadata about retrieved content.

Agents and Tools - Module 04

@Tool Annotation - Marks Java methods as AI-callable tools.

ReAct Pattern - Reason → Act → Observe → Repeat.

Session Management - Separate contexts for different users.

Tool - Function an AI agent can call.

Tool Description - Documentation of tool purpose and parameters.

Agentic Module - Module 05

@Agent Annotation - Marks interfaces as AI agents with declarative behavior definition.

Agent Listener - Hook for monitoring agent execution via beforeAgentInvocation() and afterAgentInvocation().

Agentic Scope - Shared memory where agents store outputs using outputKey for downstream agents to consume.

AgenticServices - Factory for creating agents using agentBuilder() and supervisorBuilder().

Conditional Workflow - Route based on conditions to different specialist agents.

Human-in-the-Loop - Workflow pattern adding human checkpoints for approval or content review.

langchain4j-agentic - Maven dependency for declarative agent building (experimental).

Loop Workflow - Iterate agent execution until a condition is met (e.g., quality score ≥ 0.8).

outputKey - Agent annotation parameter specifying where results are stored in Agentic Scope.

Parallel Workflow - Run multiple agents simultaneously for independent tasks.

Response Strategy - How supervisor formulates final answer: LAST, SUMMARY, or SCORED.

Sequential Workflow - Execute agents in order where output flows to the next step.

Supervisor Agent Pattern - Advanced agentic pattern where a supervisor LLM dynamically decides which sub-agents to invoke.

Model Context Protocol (MCP) - Module 05

langchain4j-mcp - Maven dependency for MCP integration in LangChain4j.

MCP - Model Context Protocol: standard for connecting AI apps to external tools. Build once, use everywhere.

MCP Client - Application that connects to MCP servers to discover and use tools.

MCP Server - Service exposing tools via MCP with clear descriptions and parameter schemas.

McpToolProvider - LangChain4j component that wraps MCP tools for use in AI services and agents.

McpTransport - Interface for MCP communication. Implementations include Stdio and HTTP.

Stdio Transport - Local process transport via stdin/stdout. Useful for filesystem access or command-line tools.

StdioMcpTransport - LangChain4j implementation spawning MCP server as subprocess.

Tool Discovery - Client queries server for available tools with descriptions and schemas.

Azure Services - Module 01

Azure AI Search - Cloud search with vector capabilities. Module 03

Azure Developer CLI (azd) - Deploys Azure resources.

Azure OpenAI - Microsoft's enterprise AI service.

Bicep - Azure infrastructure-as-code language. Infrastructure Guide

Deployment Name - Name for model deployment in Azure.

GPT-5.2 - Latest OpenAI model with reasoning control. Module 02

Testing and Development - Testing Guide

Dev Container - Containerized development environment. Configuration

In-Memory Testing - Testing with in-memory storage.

Integration Testing - Testing with real infrastructure.

Maven - Java build automation tool.

Mockito - Java mocking framework.

Spring Boot - Java application framework. Module 01