Grounded Answers: How to Check an AI Response Against Sources
By AHMAD ·
A source-checking workflow for AI assistants that summarize documents or search results.
Grounded Answers: How to Check an AI Response Against Sources
A citation is useful only if the linked source actually supports the sentence. The Microsoft RAG evaluator guidance (https://learn.microsoft.com/en-us/azure/foundry/concepts/evaluation-evaluators/rag-evaluators) treats groundedness as an answer-quality question, while Google's guidance for helpful content (https://developers.google.com/search/docs/fundamentals/creating-helpful-content) emphasizes accuracy and trust.
Make each claim checkable
Break an answer into concrete claims: a date, a feature, a comparison, or an instruction. Open the linked source and find the passage that supports each claim. A source title or search snippet is not enough. If a sentence combines two claims, it may need two sources or a rewrite.
Separate fact from interpretation
A product's documentation can support what that product says it does. It cannot by itself prove that the product is “best” for every team. Label recommendations as a judgment and explain the criteria. For changing details such as pricing, versions, and availability, record when you checked the page.
Test absence
Include a question whose answer is not in the source set. A grounded assistant should say it could not confirm the answer or ask for the missing document. Inventing an answer and attaching a nearby citation is still a failure.
Use the check in a workflow
Before publishing a guide, make a short claim ledger: claim, source URL, supporting passage, date, and status. Revisit the source if the claim changes. This takes more time than collecting links at the end, but it catches mismatches early.
The same method helps a RAG system: evaluate whether the right source was retrieved and whether the final answer stayed inside it. See our RAG evaluation guide (https://ubits.tech/learning/p/evaluate-rag-system-retrieval-answer).
Primary sources: Microsoft RAG evaluators (https://learn.microsoft.com/en-us/azure/foundry/concepts/evaluation-evaluators/rag-evaluators) · Google…