Faithfulness

Faithfulness, also called groundedness, asks whether an answer sticks to its sources. A faithful answer makes only claims that the retrieved passages, documents, or records support. An unfaithful one adds facts from the model's general knowledge, misreads a passage, or overstates what the source says, even if the result sounds reasonable.

It matters most in RAG, summarization, and any feature where users trust the output to reflect specific material. It's separate from correctness: an answer can be faithful to an outdated document and still be wrong, and a correct answer drawn from general knowledge is still unfaithful to your sources, which is a problem when you've promised answers from your docs.

Measuring it usually means breaking an answer into individual claims and checking each against the context, either with a model-graded eval using a careful rubric or with human review on a sample. Calibrate any automated grader against human labels before trusting its numbers. Separating faithfulness from retrieval quality tells you whether to fix search or fix generation.

To improve it, instruct the model to answer only from provided context, require citations or quotes, verify quotes in code, and consider higher effort or a stronger model when the sources are complex.