05 · Embeddings, RAG & Vector Search · Topic 04

What is Grounding?

Grounding connects an AI response to trusted information supplied at runtime.

What is it?

Grounding connects an AI response to trusted information supplied at runtime.

A model can generate plausible text from learned patterns. Grounding gives it specific evidence for the current task, such as documents, databases, web results, tool outputs, company records, or user-provided files.

Think of it like...

Answer with the book open.

With a source, the system can connect an answer to evidence supplied for the task. The analogy explains the role of evidence; it does not guarantee that the answer is correct.

How it works

1 · QuestionIdentify the needUnderstand what the user asks.

2 · RetrieveFind evidenceFetch trusted and relevant information.

3 · Add contextSupply the evidencePlace it into the model's context.

4 · GenerateWrite the answerUse the supplied evidence.

5 · VerifyCheck the resultShow citations or review important claims.

Real-world examples

Business example · Company policy

Input: a question about maternity leave

AI / System: retrieves the current HR document

Output: an answer based on that document

Business example · Financial data

Input: a question about last month's revenue

AI / System: retrieves current database records

Output: an answer based on retrieved data

What it is NOT

Grounding ≠ RAG

Grounding: the goal of tying answers to external evidence.

RAG: one common architecture used to achieve that goal.

Grounding ≠ Fine-tuning

Grounding: adds information at runtime.

Fine-tuning: changes model behavior through training.

Grounding ≠ Guaranteed Truth

Grounding: gives the model better evidence.

Guaranteed truth: something grounding alone cannot provide.

Remember this

Grounding ties an AI answer to trusted external information.

Independent explainer

Video

visual explainer