AI Systems Research guide

RAG vs fine-tuning vs agents: three jobs, not a ladder

A vector store is not a trained model. A trained model is not a worker. A worker without a ceiling is a demo.

Three workshop benches: a single folder under a lamp, a small kiln with a glowing door, an unrolled tool roll beside a red refuse button
Lookup. Weights. Worker. Three benches.

A chatbot can inherit an API key and still be unable to open the source behind its sentence. That failure is why RAG, fine-tuning, and agents must be separated by job.

RAG vs fine-tuning vs agents. Three jobs. Not a maturity model.

What is RAG? owns the lookup. What is an AI agent? owns the worker. This page is the comparison.

Three artefacts

RAG. Retrieval-augmented generation. Look up a named passage in a corpus you gave the system, then write. The artefact is an openable locator: file, page, chunk, date. If you cannot open it, you do not have RAG. You have extra text in a prompt. Evaluate the reasoning, not the fluency is the same test.

Fine-tuning. A weight change so a stable skill fires without fetching a document: a tone, a format, a classifier that does not move every Tuesday. The artefact is a trained checkpoint with an eval that still holds on held-out cases. Fine-tuning does not make the model know last week’s contract. A changed PDF is a retrieval problem. Treating fine-tuning as a filing cabinet is how stale clauses survive a model upgrade.

Agent. A system that holds a goal, uses tools, and continues across steps under a charter. The artefact is seven written fields and a named person who can fail the work. Using AI agents effectively is the operating loop. An agent may call retrieval as a tool. That does not turn the agent into RAG, and it does not replace the lookup test.

Not a ladder

Doing all three is allowed. Doing them in a fashionable order is not a strategy.

Fail the lookup: you need RAG, or a refuse, not a bigger model.

Fail a stable format that never needs a document: fine-tuning can be the job. A weekly handbook is not that job.

Fail the ceiling: you need a charter, not another index.

A vector store is furniture. GRAIN, on this site, is the inspectable retrieval loop. Fine-tuning is not GRAIN. An agent without Inspect still answers from the closest wrong paragraph.

A proposal you can fail

If a vendor sells one retainer for all three, ask which artefact you will hold on Friday: a locator you can open, a checkpoint with an eval, or a charter with a named reviewer. Missing all three, you bought a demo.

The RAG Engineer is live. It does not turn fine-tuning into a source of truth. Write which job failed. Then stop stacking rungs.

Terminology

Sources

  1. What is RAG?
  2. What is an AI agent?
  3. Using AI agents effectively
  4. Evaluate the reasoning, not the fluency
  5. The RAG Engineer

Further reading

Markdown for LLMs