28/02/2026
AI is moving fast. Structure helps. So we have an AI Periodic Table to organize the ecosystem into something more systematic, understandable, and usable — from Prompts, Embeddings, & LLMs to RAG, Agents, Guard rails, Red-Teaming, Multi-Agent systems, and Thinking models. The goal is simple: make AI concepts easier to map, learn, discuss, and build with. Because when the landscape gets noisy, frameworks create clarity.Inspired by the idea of turning complex AI architecture into visual system for builders, strategists, & learners.
R1 — Primitive
(1) Pr — Prompts
The instruction layer of AI. Prompts tell a model what to do, how to behave, what context to use, and what format to return. They are the most direct human-to-model control interface. Prompting is treated as one of the most basic building blocks in the referenced explainer.
(2) Em — Embeddings
Embeddings are numerical representations of meaning. They convert text, images, or other data into vectors so similarity can be measured. They power semantic search, clustering, recommendation, and retrieval pipelines. The explainer describes embeddings as core to semantic search and vector databases.
(3) Lg — LLMs
Large Language Models are the foundation models that generate, summarize, reason, classify, and converse in natural language. They serve as the core “brain” for many AI systems and are a stable foundational capability in the referenced overview.
R2 — Compositions
(1) Fc — Function Cell / Function Calling
This is the action bridge between a model and external tools. Through function calling, an LLM can trigger APIs, calculators, databases, CRMs, search tools, or workflows instead of only generating text. In the explainer, this is presented as a key composition layer.
(2) Vx — Vector
This likely refers to vector storage or vector databases. These systems store embeddings and make retrieval possible using similarity search. They are essential for semantic search, memory-like recall, and RAG pipelines.
(3) Rg — RAG
Retrieval-Augmented Generation combines search/retrieval with generation. Instead of relying only on what a model “knows,” RAG fetches relevant external information first, then uses it to answer more accurately. It helps with freshness, grounding, and enterprise knowledge access. The IBM-oriented summary highlights grounding AI responses in organizational knowledge as a major architectural layer.
(4) Gr — Guardrails
Guardrails are the control and safety mechanisms around AI systems. They include input filtering, output moderation, policy checks, grounding rules, permission limits, and workflow constraints. They reduce hallucinations, unsafe responses, and business risk. Governance and control layers are emphasized in the IBM-style breakdown.
(5) Mm — Multi Model
This refers to systems that use more than one model together. For example, one model for OCR, another for retrieval, another for generation, and another for ranking. Multi-model setups are common in production AI because different tasks are best handled by specialized models.
R3 — Deployment
(1) Ag — Agent
Agents are AI systems that can plan, decide, act, observe results, and iterate toward a goal. They often combine LLMs, tools, memory, and workflows. The referenced explainer describes agents as moving toward autonomy through think-act-observe loops.
(2) Ft — Finetune
Fine-tuning means adapting a base model using domain-specific data so it behaves better for a specific task, tone, or knowledge pattern. This is useful for specialized domains like legal, medical, education, or enterprise workflows. The explainer identifies fine-tuning as adapting the model to specific data or use cases.
(3) Fw — Framework
Frameworks are the development and orchestration environments used to build AI systems. Examples include agent frameworks, prompt orchestration tools, RAG frameworks, evaluation frameworks, and workflow engines. They provide reusable structure, speed, and consistency.
(4) Re — Red-Team
Red-teaming is the deliberate stress-testing of AI systems to uncover weaknesses. This includes testing for hallucinations, prompt injection, jailbreaks, toxic output, bias, data leakage, and unsafe tool usage. It is essential for production readiness and responsible deployment.
(5) Sm — Small
This likely refers to small models or lightweight models. These are cheaper, faster, and easier to deploy than large frontier models. Small models are useful for edge devices, narrow tasks, classification, and cost-sensitive deployments.
R4 — Emerging
(1) Ma — Multi-Agent
Multi-agent systems involve several AI agents working together, either collaboratively or through role-based coordination. One agent may research, another may plan, another may verify, and another may execute. This pattern is increasingly used for complex workflows.
(2) Sy — Synthetic
Synthetic usually refers to synthetic data, synthetic content, or synthetic environments. In AI, this can mean generated datasets for training/testing, simulated conversations, synthetic customer scenarios, or artificial examples used to improve models and evaluations.
(3) In — Interpret
Interpretability focuses on understanding why an AI system produced a result. It includes transparency, explainability, attribution, confidence signals, and model behavior analysis. This becomes more important in regulated or high-stakes sectors.
(4) Th — Thinking
This likely refers to reasoning-oriented models or systems designed for deeper stepwise problem-solving, planning, and reflective inference. In practical terms, it points toward models that don’t just generate language but handle structured reasoning more effectively.
