October workshops are openBuild a Search AI Agent$99 early bird

Enroll now
Skip to content

Lesson 4 of 7 · About 3 minutes

Make the answer earn your trust

Match a claim to a relevant source and spot what remains unknown.

01 / Learn the idea

Evidence before confidence

A model's learned knowledge is not a live copy of your school's documents. Retrieval finds relevant material to use when generating an answer. Retrieval-augmented generation, or RAG, combines this material with a language model. The original RAG paper explored this approach using a searchable collection of Wikipedia passages.[1]

Grounding connects a response to supporting information. Citations make that connection inspectable: you can open the source and compare it with the claim. Check who produced the source, whether it applies to this question, and whether the cited passage actually supports the answer. A link beside a sentence is not proof by itself.[2]

Suppose last year's club notice says Friday, while this year's signed notice says Thursday. The older document might be real yet irrelevant to this year's meeting. If neither document gives a room number, the honest answer leaves the room unknown and asks for the missing information.

Original school example

A missing room

Source: the current school notice names Thursday at 4 pm but no room. Supported answer: Thursday at 4 pm. Unsupported addition: Room 204. Search for the room separately or say it is not established.

A common misconception: “An answer with citations cannot hallucinate.

The source may be outdated, irrelevant, or misread. Check the specific claim against the actual passage.

02 / Explore the flow · Guided simulation

A real passage can still be the wrong evidence

A fictional school policy changes on September 1. Find the rule for a September 15 event, then check whether the required approval is present.

Original fictional scenario. Decisions are authored; calculations and checks run in your browser. No live AI or external action.

1Goal
2Action
3Observation
4Check
5Stop / ask

↳ A failed check can send the flow back to another action—or to a human.

Step 1 of 4 · Visible system event

Question

What does the fictional policy require for the September 15 event?

Inspect the visible trace (1 events)
  1. Question

    What does the fictional policy require for the September 15 event?

This log shows authored actions, tool results, and checks. It is not a model’s private reasoning.

03 / Check your understanding

A current notice confirms a club meeting on Thursday but says nothing about the room. Which answer is supported?

Choose your answer

Where these ideas come from

Primary sources, explained in our own words. Research findings and engineering guidance are different kinds of evidence; neither is a guarantee that an agent will succeed.

[1] Facebook AI Research, UCL, and NYU · Research paper · 2020-05-22Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (opens in a new tab)

Supports: Combining retrieved documents with text generation.

Limit: NeurIPS 2020 research on particular models and datasets. Retrieval is not a guarantee of factual correctness.

[2] Google · Technical documentationGrounding with Google Search (opens in a new tab)

Supports: Connecting answer passages to source citations.

Limit: Explains source attribution in a product. A citation still needs checking against the claim it accompanies.

Stemtiq original lesson · Sources checked September 12, 2026 · Editorial review pending. School examples and guided simulations are original simplifications, not replications of the cited systems. No research affiliation or endorsement is implied.