Skip to main content
Explainers · Grounding & RAG
A Visual Primer

From memory
to evidence.

The single most effective way to make AI reliable: stop asking what it remembers and start showing it the documents. Six stages.

Scroll
01

The problem with memory.

A language model’s knowledge was frozen on the day its training ended. And even within that, recall is reconstruction — patterns, not records. Ask it about your specific contract, your specific witness, your specific deadline, and it has nothing to recall — so it may produce something plausible instead.

What it was trained on
The general shape of the law — millions of pages of statutes, judgments and commentary, blended into pattern.
→
The gap where invention happens
What you need
Clause 14 of THIS contract — the actual words, in the actual document, in front of you now.
Between the general and the specific sits the gap — and the model fills gaps with plausibility.
02

Show, don’t ask.

Grounding means putting the actual documents into the conversation and instructing the model to answer only from them. The question changes from “what do you remember about limitation periods?” to “what does this contract say about limitation?” — a task language models are genuinely excellent at.

UNGROUNDEDAsk: What does the case law say about restrictive covenants?

The model answers from pattern — with no document to check it against.
→
GROUNDED<covenant>
[the clause text, pasted in full]
</covenant>

<task>
Set out what this clause restricts, quoting the words you rely on.
</task>
03

The same question, twice.

“When does the notice period under this agreement expire?”
ANSWER — FROM MEMORYNotice periods under commercial agreements of this kind are typically 30 days from service. Assuming notice was served in mid-July, the period would therefore likely expire on 14 August.
Plausible. Generic. Unverifiable — the model has never seen the agreement.
Same question. Same model. The document is the difference.
04

RAG, in one diagram.

Retrieval-Augmented Generation is grounding at scale. When the documents are too many to paste — a disclosure exercise, a precedent bank, an entire bundle — a search step finds the relevant passages and inserts them into the context automatically.

1 · Your question
Asked in plain language, as you would ask a junior.
→
2 · Search your documents
The relevant passages are retrieved from your own materials — not the open internet.
→
3 · Model answers
From the retrieved passages, with references back to where each answer came from.
When a legal tool says it “searches your knowledge base” or “cites its sources”, this is usually what’s happening underneath.
05

What it fixes. What it doesn’t.

✓ GROUNDING FIXES
  • Recall. Answers come from the page, not the model’s memory.
  • Currency. Your documents can be newer than the model’s training.
  • Checkability. Quotes give you something to verify against the source.
✕ GROUNDING DOESN’T FIX
  • Retrieval misses. The search can miss the key passage — the model answers from what was found, not what exists.
  • Reasoning errors. Correct sources can still be misapplied on top.
  • Overstatement. A model can still overstate what a passage supports.
Grounding narrows the space for invention — it does not abolish the need to read the source.
06

Source-locked, by instruction.

The practical pattern: tell the model it may use only the supplied documents, must quote what it relies on, and must say “not in the materials” rather than fill gaps.

Answer only from the documents provided. For every proposition, quote the passage you rely on with its reference. If the documents do not answer the question, say “not in the materials” — do not infer, do not fill gaps.
This one instruction converts the model from a raconteur into a witness under discipline.

Evidence in, evidence out.

Grounding is the bridge between a model that sounds right and one you can check. Next in the series: how to brief the model precisely once it has the documents.

Next: Prompting Opus 4.8 for Legal Work →