TerminologytermStep 4: Grounding & specializingAllLegalOperations

Term: Retrieval-Augmented Generation (RAG)

~7 min read

Estimated time: ~7 min read — for the in-app brief plus opening the primary source.

What this is

RAG connects a language model to your approved documents or systems so answers can be grounded in your content rather than only general training.

Everyday example

A 'chat with our policies' bot that searches your approved SharePoint or contract repository, then drafts an answer with citations, is usually RAG — not the model magically memorizing your files.

RAG retrieves your approved documents first, then asks the model to answer from that material.

  • It is how you get answers about your policy, not a generic internet guess.
  • It fails if retrieval is poor — wrong chunk, stale file, missing access control.
  • It is usually cheaper and more controllable than fine-tuning for “know our files.”
  • Still requires a human on high-stakes answers.

Next action: If a bot must know your policy, ask whether it retrieves a named corpus — and who updates that corpus.

What changes in how you lead

How decision rights, process, and ownership should change.

  • Knowledge owners (Legal, HR, Product) own the corpus, not only IT.
  • Access control on retrieval is as important as access control on the drive.

Compare related ideas

Retrieval-Augmented Generation (RAG) vs Fine-tuning

RAG looks up your documents at answer time (good for changing policies and citations). Fine-tuning reshapes how the model behaves from examples (style, format, specialized patterns) but does not replace a live document library.

Open Fine-tuning

Retrieval-Augmented Generation (RAG) vs Pasting files into ChatGPT / Claude / Grok

Manual paste is a one-off context dump. RAG is a designed system with search, permissions, and usually citations — better for enterprise scale.

Deep dive

Helps reduce some hallucinations when retrieval is high quality — not a guarantee of truth.

Ask: which repositories, who can access, how often indexed, how citations work?

Permissions must match source systems; a chatbot should not leak restricted files.

Related terms

Related weekly lessons

terminologyragenterprise