FOUNDATIONS & DEFINITIONS

AI Engineering & Systems Glossary

Essential reference guide for software architects, systems engineers, and technical founders. Precise technical definitions and architectural context for RAG pipelines, autonomous agents, n8n workflows, and low-latency voice AI systems.

Autonomous SystemsAgentic AI

AI Agent

An autonomous software system driven by foundation models that perceives environment state, plans multi-step execution graphs, executes tool calls, and dynamically adjusts its strategy.

Unlike simple one-shot chatbots, production AI agents maintain execution state machines, enforce permission boundaries, and use human-in-the-loop checkpoints for high-stakes operational tasks.

Autonomous SystemsTool Calling

Function & Tool Calling

The structured mechanism enabling an LLM to generate schema-compliant JSON payloads that invoke external APIs, execute SQL queries, or trigger operational scripts.

Production tool execution requires strict schema validation (via Zod or Pydantic), deterministic timeouts, execution sandboxes, and human-in-the-loop approval gates for destructive operations.

Protocols & StandardsMCP

Model Context Protocol

An open protocol standardizing how AI applications and autonomous agents interface with external tools, repositories, and context providers over uniform JSON-RPC transports.

MCP decouples client agents from proprietary tool calling adapters, enabling modular, vendor-agnostic tool servers that can be reused across Claude, OpenAI, and local LLMs.

Security & SafetyInjection Attack

Prompt Injection

A critical AI security vulnerability where untrusted user input or malicious third-party document content overrides the original system prompt instructions, hijacking the agent's behavior.

Defense-in-depth against prompt injection requires dual LLM verification checks, input boundary sanitization, read-only database connections, and hardened tool permission scopes.

Retrieval & KnowledgeReranker

Reranking & Cross-Encoders

A second-stage retrieval optimization where candidate context chunks retrieved via initial bi-encoder vector search are scored for precise semantic relevance using a heavy cross-encoder model.

Rerankers evaluate query-document pairs simultaneously, correcting vector distance bias and ensuring the top 3-5 chunks fed to the LLM context window contain the highest information density.

Retrieval & KnowledgeRAG

Retrieval-Augmented Generation

An AI architecture that enhances large language model responses by retrieving verified, authoritative context from external databases before generating an answer.

Rather than relying solely on frozen parametric weights, RAG combines semantic vector search and lexical BM25 matching to inject precise ground-truth chunks into LLM system prompts, minimizing hallucination.

Cost & PerformanceSemantic Cache

Semantic Caching

An infrastructure optimization pattern that stores prior prompt-response pairs in an in-memory vector cache (like Redis), serving cached completions when query semantic similarity exceeds a set threshold.

Semantic caching drastically reduces API token expenditures and reduces endpoint response times from seconds to sub-20ms for recurring user questions.

Retrieval & KnowledgeVector DB

Vector Database

A database engineered to store, index, and query billions of high-dimensional vectors with sub-100ms latency using Approximate Nearest Neighbor (ANN) indexing structures.

Common vector storage engines include dedicated stores like Pinecone, Milvus, and Qdrant, alongside relational extensions like pgvector and Oracle AI Vector Search.

Retrieval & KnowledgeEmbeddings

Vector Embeddings

Dense mathematical vector arrays representing words, sentences, or code snippets, positioned so that geometric proximity corresponds directly to semantic similarity.

Generated by specialized embedding models (such as text-embedding-3-small or BGE-large), embeddings power cosine similarity and dot-product vector search across enterprise document corpora.

Workflow Automationn8n

n8n Automation

An open-source, source-available workflow automation platform that enables deterministic node-based orchestration across APIs, webhooks, databases, and generative AI agents.

In enterprise environments, n8n scales via Redis queue mode with dedicated worker pools, self-healing retry strategies, and automated dead-letter alerting.