AI Grounds

Open AI Grounds on a desktop

These interactive lessons need a larger screen. Please continue on a desktop or laptop computer.

Guided discovery

From Retrieval to a Grounded Answer

Change the context. Trace the answer evidence.

Your fictional corpus · Harbor Museum

Authored rankings and answer bank · no live retrieval, embeddings or LLM. Four fixed documents and three queries make each failure stage inspectable.

Retrieve the first N ranked candidates; the full candidate list remains visible. Use arrow keys on the slider. Press Enter or leave the number field to apply an exact edit.

Supply up to this many retrieved chunks to the answer bank. It cannot add unretrieved documents. Use arrow keys on the slider. Press Enter or leave the number field to apply an exact edit.

1. Ranked candidates

2 retrieved of four candidates.

1. D1 · Current hours

Tuesday opening: 10:00.

2026-01-01 · Retrieved

2. D2 · Old hours

Tuesday opening: 09:00. Outdated.

2022-08-15 · Retrieved · Outdated

3. D3 · Admission

Admission: 4 credits.

2026-01-01 · Outside retrieval budget

4. D4 · Weekly closure

The museum is closed on Sunday.

2026-01-01 · Outside retrieval budget

2. Supplied context

2 supplied · 2 relevant to this query.

D1 · Current hours

Tuesday opening: 10:00.

D2 · Old hours

Tuesday opening: 09:00. Outdated.

3. Authored answer

Tuesday at 10:00. [D1]

Source support
Supported
Requested field filled
Yes
Matches full current reference
Yes

2 retrieved → 2 supplied → source-supported claim; requested field filled; full reference matches.

Current reference and exact support ledger

Current reference: Tuesday at10:00. Source support is checked against a cited supplied document's authored facts. Requested-field coverage means supplying an amount/time/day, even if wrong. Full-reference match checks the requested fact; an incomplete answer can contain a true partial claim while failing that match.

DocumentRelevant?Supplied?Contains answer fact?
D1YesYesYes
D2YesYesNo
D3NoNoNo
D4NoNoNo
Authored ordering and answer rules

Relevant first: Hours D1,D2,D3,D4; Admission D3,D1,D2,D4; Closure D4,D1,D2,D3. Outdated first always uses D2,D3,D4,D1. These rankings are a lookup table, not computed embeddings. The candidate list includes rows outside the retrieval budget; only Retrieved rows can enter supplied context.

Use evidence prefers a relevant current document, otherwise a supplied relevant old one; with none, it abstains. Wrong claim produces11:00,7 credits or Monday while citing the available relevant source. Incomplete says Tuesday, that a charge exists, or that a weekly closure exists; it omits the requested time/amount/day. This is a disclosed answer bank, not a model guaranteed to behave this way. Context budget is a chunk count, not a token limit. Current reference is an authored independent ledger. Real RAG must assess retrieval and generated claims with real data and evaluation.

Retrieval-Augmented Generation · Original paper