ADR (Architecture Decision Record)

A short record of an architecture decision: context, decision, alternatives and trade-offs. Makes the why behind choices traceable.

→ article

AI agent

A system where a model decides the next steps and calls tools to reach a goal. In production, it is a distributed system with autonomy.

→ article→ pattern

AI FinOps

The practice of measuring and optimizing AI cost per completed task, not per call, using caching, routing and limits.

→ article→ pattern

AI observability

The ability to explain what an AI system did and why: context used, versions, decisions, cost and quality, tied together by a trace.

→ article→ pattern

Anti-corruption layer

A layer that translates and protects a legacy system, exposing a clean domain model to consumers, including AI.

→ article→ pattern

Context engineering

The discipline of designing what goes into the model context: sources, format, order, size and permissions. It goes beyond writing good prompts.

→ article

Context orchestration

The layer that decides where information comes from, how it is transformed, who may use it and how it reaches the model, in a governed, reusable way.

→ article→ pattern

Context window

The maximum number of tokens the model considers in one call. Fitting in the window does not mean everything should be there.

Embedding

A numeric representation of text that lets you compare meanings. The basis of semantic search in many RAG systems.

Evaluation (eval)

A set of cases and metrics that measures the quality of an AI system. Without evaluation, prompt or model changes are guesswork.

→ article→ pattern

Grounding

Anchoring the model answer in verifiable sources provided in context, with citations, instead of relying only on what it learned.

Guardrails

Controls outside the model that limit inputs, outputs and actions: validation, filters, permissions and approvals. A prompt is not a guardrail.

→ article

Hallucination

A fluent, incorrect answer. Most of the time it is a symptom of poor or missing context, and is reduced with architecture, not just the model.

→ article

Human in the loop

A step in the flow where a person approves or corrects a system action before it happens, usually for irreversible actions.

→ pattern

MCP (Model Context Protocol)

An open protocol that standardizes how models access data sources and tools. Enables reusable integrations and central governance.

→ article→ pattern

Prompt caching

Reusing the processing of stable prompt parts across calls. Cuts cost and latency when the prefix does not change.

→ pattern

Prompt injection

Malicious instructions hidden in content the model reads (documents, emails, pages) trying to hijack its behavior. Mitigated with limits outside the model.

RAG (Retrieval-Augmented Generation)

A pattern where the system retrieves relevant information and gives it to the model before answering. Quality depends more on architecture than on the model.

→ article→ pattern

Token

The unit of text a model processes and bills. Price per token is only part of the real cost of AI in production.

→ article

Trace

A chained record of every step of an interaction, from request to answer, under a single identifier. The unit of truth in observability.

→ pattern