Artificial Intelligence · Frontend

The AI Add-On Map: LLM, RAG, MCP, Agent, and Skill, All Plotted on One Diagram

By 尘世中一位迷途小书童 · Oct 30, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The cost of fuzzy AI vocabulary is concrete: you reach for the wrong tool and get plausible fiction instead of grounded results. Knowing which add-on fills which hole determines whether a chat box becomes a reliable assistant or a polished hallucination engine.

Summary

Most AI vocabulary confusion comes from treating each new term as a separate 'smarter model'. These terms are add-ons to a single language model, and one concrete task keeps them straight: 'Based on the latest login rules, update the API documentation in the repo, self-test, then show me.' Each term answers the same question: what does AI lack here, and what does this term fill? The result is a layered map from the model's language ability outward to retrieval, tool execution, multi-step loops, standard connectors, and reusable workflow cards.

The walkthrough covers Prompt, context window, and token as the interaction layer; Embedding and vector DB as meaning-based search; RAG as open-book grounding; tool calling as borrowed hands; Agent as a multi-step decision loop; MCP as a standard plug for external tools; and Skill as a copy of the house rules. Common confusions are also addressed, such as mistaking hallucination for a sign of a broken model, or treating fine-tuning as a cheap fix when rules change often.

A zero-dependency demo shows the practical difference: without retrieval, the model answers confidently with plausible but wrong rules ('lock 24 hours after 3 tries, MD5'), with a simple keyword retrieval it cites the correct clauses ('lock 15 minutes after 5 tries, Argon2id'). The comparison table makes each concept's missing piece explicit, and the full chain shows how Skill, Agent, RAG, MCP, tools, and a final human approval fit together.

Takeaways

— RAG is retrieval-augmented generation: it retrieves relevant documents, stuffs them into the context, then generates an answer.

— Tool calling lets the model propose actions, but the runtime executes them; the model does not grow hands.

— An Agent is a multi-step decision loop, not just a chat bubble with a model behind it.

— MCP is an open protocol from Anthropic that standardizes how data sources and tools connect to AI applications.

— Skill is a reusable workflow card that encodes steps and constraints; it is not model fine-tuning.

— Hallucination is a normal risk when the model has no external grounding, not a sign of a broken model.

— Context windows are finite and are not permanent memory; tokens are a billing and length unit.

— Fine-tuning is expensive and slow; for frequently changing rules, RAG plus Skill is the better default.

— The demo contrasts a fake answer ('lock 24h after 3 tries, MD5') with a retrieved answer ('lock 15min after 5 tries, Argon2id').

— The full chain: Skill sets rules, Agent breaks steps, RAG retrieves, tools act, you approve.

Conclusions

The framing of all these technologies as add-ons to a single language model is a practical antidote to the tendency to treat each new acronym as a separate 'AI'.

The concrete use case makes the distinction between RAG (grounding) and Agent (action) tangible: RAG decides what to say, tools decide what to do.

The claim that fine-tuning is overkill for frequently changing rules and that RAG plus Skill is the pragmatic default is a useful counterweight to the hype around model customization.

Presenting Multi-Agent not as an automatic win but as a source of orchestration overhead keeps expectations realistic.

Concepts & terms

RAG

Retrieval-Augmented Generation: retrieves relevant documents and inserts them into the context before the model generates an answer, giving it a basis for responding.

MCP

Model Context Protocol: an open protocol from Anthropic that standardizes how AI applications discover and call external data sources and tools.

Skill

A reusable workflow card that encodes the steps, constraints, and output format for a specific task, so the assistant follows your company's rules instead of improvising.

Agent

A loop in which the model repeatedly decides the next action, calls tools or retrieves information, observes the result, and continues until the goal is reached.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗