← back to the archiveCover illustration for “Groundedness: is the AI's answer backed by its sources?”
TERM7mo ago·Published · Updated ·by Andy Padia

Groundedness: is the AI's answer backed by its sources?

In short: Groundedness, also called faithfulness, measures whether every claim in an AI answer is supported by the sources the system was actually given.

Groundedness, also called faithfulness, measures whether every claim in an AI system's answer is supported by the sources it was given, such as retrieved documents or tool results. A grounded answer may be incomplete, but it does not add facts from nowhere. The Ragas paper of September 2023 proposed scoring RAG pipelines this way without ground-truth human annotations.

Groundedness in plain words

Every sentence in the answer should be able to point at the line in the documents it came from. If it cannot, it is opinion or invention.

How it works

The answer is split into individual claims. Each claim is checked against the provided context by an LLM judge, a natural-language inference model or a grounding service, and marked supported or unsupported. A common score is the share of supported claims. Some systems also check that citations point to the right passage.

Google's check grounding API, as documented in November 2026, works this way. It treats each sentence as one claim, returns a support score from 0 to 1 with citations to the supporting facts, and counts a claim as grounded only when the facts support it fully. A sentence with the right names and the wrong date is ungrounded.

Example

As an illustration, the sources state "Refunds are available within 30 days of purchase." The answer says "You can get a refund within 30 days, including on discounted items." The first claim is grounded; "including on discounted items" is not in the source, so the answer scores below full marks and gets flagged.

Where you'll meet it

  • RAG evaluation libraries such as Ragas
  • Cloud provider grounding-check APIs
  • Observability tools scoring production answers
  • Regulated domains where unsupported statements carry risk

What it changes for you

For developers, groundedness separates two failures: retrieval missed the facts, or the model ignored or embellished them. Each needs a different fix. For product teams, low-groundedness answers can be blocked, rewritten or routed to a person. For compliance, it is evidence that answers stay within approved material.

Groundedness vs correctness

Groundedness asks "is it supported by the sources given?" Correctness asks "is it true?" A grounded answer built on an outdated document can still be wrong.

Common mistake

My view: measuring groundedness and ignoring relevance. An answer can be perfectly grounded in the wrong passage and still useless. Score both whether the answer is supported and whether it addresses the question.

Hallucination · Retrieval-augmented generation · LLM-as-a-judge · Evals · Reranking

From the archive: Why your RAG pipeline returns irrelevant results

Score whether the answer is supported and whether it addresses the question, because one without the other misleads.

#field-guide#evaluation#rag#reliability

Sources

  1. arxiv.org/abs/2309.15217
  2. docs.cloud.google.com/generative-ai-app-builder/docs/check-grounding
← older drop
Claude Code hooks: shell commands at agent lifecycle points
newer drop →
Model routing: sending each request to the right model

related drops

explore all 557 drops →
← back to the archiveday 175