Grounding
Also: 有根據 · 接地 · 可溯源 · grounded · 可查證
Tying each claim the model makes back to a source the reader can verify themselves; not "it says it has a source" but "the source really exists and really supports this sentence".
When you will meet it
This is what pulls AI from "sounds right" to "verifiably right". In enterprise, finance and healthcare — where being wrong carries liability — an unverifiable answer is as good as no answer. Understanding how grounding differs from "just attach a link" tells you which AI outputs you dare to use.
An analogy
Like footnotes in a paper. A bad footnote is a bibliography at the end listing titles that correspond to nothing in the text (looks cited); true grounding marks each claim with a number, and when you flip to that page, the sentence genuinely supports it. The former is decoration; the latter is verifiable.
Minimal example
主張：「本公司 2024 年營收增長 18%」
✗ 只是「引用來源」：
回答末尾附上「資料來源：年報」
→ 但年報有 200 頁，18% 這個數從哪來？沒法查
✓ 真正「有 grounding」：
「營收增長 18%〔年報第 45 頁，合併損益表〕」
→ 翻到第 45 頁，那個數字確實在、確實支持這句
差別：能不能被一個讀者在合理時間內，自己核對到The key is the last line: the standard for grounding is not "did it attach a source" but "can a reader actually follow the source and verify that sentence". Attaching a vague attribution is easy but is not grounding; precisely locating a claim to a checkable page, table or passage is. This is also RAG's core advantage over a bare model: it inherently carries "this sentence came from this retrieved passage".
What people get wrong
- Treating "the model says it has a source" as "it is grounded". The model will fabricate the source too (hallucination); the page numbers and links it claims can be fake. The test for grounding is that you actually check the source, not that it claims one.
- Assuming RAG makes you grounded automatically. RAG provides the channel for "which passage a claim may come from", but if the system does not carry and show the retrieved source to the user, the answer is still unverifiable. Grounding requires carrying the source all the way to the final output.