AI Glossary: Prompt, RAG, Fine-Tuning, and Agents

Introduction: AI Terms Get Much Easier Once You Cross the "Katakana-English Wall"

If you follow generative AI topics, technical terms keep appearing one after another—prompt, RAG, fine-tuning, agent... They are hard to approach at first, but in fact they are all words of role division that express "how you have AI work for you."

In this article, we organize the important words frequently used in the field as accessibly as possible, in the order "in a word," "when do you use it?", "merits/cautions," and "how to choose."

Prompt: A "Request Message" to AI

In a word

A prompt is the instruction text that tells generative AI "what you want done and how." It is the input for the AI and an important point that determines the quality of the deliverable.

When do you use it?

• Writing (emails, proposals, articles, summaries)
• Code generation (creating functions, fixing bugs, review)
• Image generation (composition/style specification)
• Data organization (formatting tables, classification, extraction)

Tip: prompts get stronger with "role," "constraints," and "examples"

The three classic templates that work well are:

1. Role: please answer as an expert in XX
2. Constraints: character count, tone, output format (JSON/table/bullets)
3. Examples: good examples, bad examples, an expected output sample

Example: "You are a PM for BtoB SaaS. Produce 3 new-feature ideas, each with aim, target user, and KPI, in bullets. Briefly supplement technical terms."

Cautions

• Vague instructions tend to make answers vague too (the "make it nice" problem)
• It diverges with insufficient shared premises (target, purpose, prohibitions)
• Also be careful about information management, such as not including confidential information

RAG (Retrieval-Augmented Generation): A Mechanism to Add "Basis" Through Search

In a word

RAG is a mechanism where, before the AI builds an answer, it searches related information from internal documents or a database (Retrieval) and then generates (Generation) based on that content. Roughly, picture "letting the AI crib from internal knowledge."

When do you use it?

• Chatbots using internal rules, procedure manuals, FAQs, and minutes
• Support answers while referring to product manuals
• Drafting sales materials and proposals (referring to past cases)
• First-line legal/compliance checks (with internal-rule citations)

Merits

• Can reflect the latest information and company-specific information (OK even if the model hasn't trained on it)
• Easier to reduce hallucination (plausible errors)
• Showing the citation source raises accountability

The basic structure of RAG (roughly)

1. Split documents (chunking) and store them
2. Vectorize documents into a searchable form (Embeddings)
3. Search for document fragments close to the question (vector search)
4. Put the search results into the prompt and generate the answer

In implementation, a vector DB (e.g., Pinecone, Weaviate, Milvus, pgvector) or a framework (e.g., LangChain, LlamaIndex) is often used.

Caution: "add RAG and everything is solved" is false

• If search is off, the AI cleverly errs on the off basis
• Document updating, permission management, confidentiality classification—operational design is important
• Accuracy changes with chunk size and metadata design (e.g., department, product name, revision)

Fine-Tuning: Having the Model Learn "Habits"

In a word

Fine-tuning is a method of retraining an existing large language model (LLM) with additional data to optimize it for a specific task or writing style. It is considered when you want to stably produce "this company's way of writing" or "this job's judgment criteria."

When do you use it?

• Want to strictly fix the answer format (e.g., always JSON, always steps 1–5)
• Want to unify phrasing/terms in a specific domain (e.g., medicine, finance, manufacturing)
• Want to make tasks like classification/extraction highly accurate
• Want to reproduce "the internal way of writing" (tone, wording)

Merits

• Reduces prompt dependence and leans toward stable output
• For frequent tasks, prompts can sometimes be shortened (which can affect speed and cost as a result)

Caution: the division with RAG matters

A common misunderstanding is "why not just fine-tune all internal documents?" In reality the following tendencies exist.

• Latest information / frequently updated information: suited to RAG (easy to update)
• Output form / judgment habits: suited to fine-tuning (stabilization)

Also, if training-data quality is low, the model learns that as-is. It's worth holding that data preparation (label consistency, removing errors) tends to be the center of the cost.

Agent: Leaving Even the "Sequencing" to AI

In a word

An agent is a mechanism where AI not only answers once but thinks through procedures to achieve a goal, uses tools, and repeats execution if needed. In human terms, it is close to a "worker" or "secretary."

When do you use it?

• Want to do research then summarization then materials in one go
• Want to automate operating internal systems for ticket creation, data retrieval, notifications, etc.
• In customer support, want to do situation check then knowledge reference then answer drafting then CRM registration

Elements that support agents

• Tool use / function calling: search, DB lookup, calendar, sending email, etc.
• Memory: retaining conversation history and user settings
• Planning: decomposing tasks and executing in order
• Guardrails: permissions, confirmation flows, execution limits (e.g., don't send arbitrarily)

Caution: as convenient as it is, accidents happen easily

Because an agent can "execute," design mistakes cause trouble—e.g., sending email to the wrong address, accessing out-of-permission data, automatically registering misinformation.

In practice, basics like inserting human approval (Human-in-the-loop) for important operations, keeping execution logs, and handing over tools with least privilege are effective.

Related Terms Together: Learn These Too and Understanding Speeds Up

LLM (Large Language Model)

A "model good at producing text" learned from large amounts of text. The central player of generative AI.

Token

The unit by which AI processes text. In Japanese it is not "1 character = 1 token"; it varies by how it is split. Pricing and input limits are often determined by tokens.

Hallucination

The phenomenon of AI plausibly lying. It is easier to suppress by giving a basis with RAG, adding constraints like "say unknown if unknown," or requiring a citation source.

Embeddings

Technology that converts text into vectors (sequences of numbers) so it can be searched by "closeness of meaning." It is the core of the search part of RAG.

Context window

The upper limit of the amount of information a model can read at once. Since you can't put long text in as-is, RAG chunk design and summarization become effective.

A Practical Guide to Usage Division: When Unsure, Think in These 3 Choices

Field judgment becomes clearer if you think in the following order.

1. First, solve with a prompt: cheap, fast, testable
2. If internal information is needed, RAG: handles knowledge that assumes updating
3. If you want to stabilize output, fine-tuning: settle the form and habits

And if you "want to execute over multiple steps," considering an agent last makes over-engineering less likely.

Conclusion: Once You Understand the Terms, AI Use Becomes a "Conversation About Design"

Prompts are "how to convey," RAG is "how to give a basis," fine-tuning is "habit forming," and agents are "sequencing and execution." Once these four differences click, AI-adoption discussions shift from "vaguely AI" toward "choosing a mechanism that fits the purpose."

Next time an AI initiative comes up internally, please ask: "Is that a story that needs RAG? Or can it go with a prompt?" The resolution of the conversation should rise sharply.