Definition
Faithfulness is the degree to which claims in a generated answer are supported by the evidence provided to the model. In a RAG system, that evidence is usually the final retrieved context passed into generation.
Faithfulness asks whether the answer stayed grounded in its supplied sources. It does not by itself ask whether those sources are correct or applicable.
Simple example
The supplied policy says damaged deliveries may be replaced within 14 days. An answer that states a 30-day replacement window is unfaithful because the evidence does not support that claim. An answer that accurately repeats the 14-day rule can be faithful even if the application retrieved an obsolete policy.
Why it matters
Models can add plausible details that are absent from the supplied evidence. A faithfulness check identifies those unsupported claims. Reviewers can ask which passage supports each statement, then investigate why any unsupported claims appeared.
One important nuance
Faithfulness is not the same as correctness. An answer can be faithful to outdated, incomplete, or unauthorized context and still be wrong for the user. Evaluate source quality, applicability, retrieval completeness, and access control separately. When using an automated evaluator, pass the exact context seen by the generation model. A larger source document can make unsupported claims look grounded even when the relevant text never reached the model.