A good fit for
- → Document assistants
- → Retrieval experiments
- → Comparing ingestion and ranking pipelines
Know the limits
- → A faithfulness score does not prove the source is true.
- → Synthetic test questions can overfit the corpus.
- → Framework defaults should not decide document permissions.
Your first useful workflow
Build a labeled set
Pair representative questions with the source passages that should support the answer.
Test retrieval alone
Check whether the required evidence appears before asking a model to answer.
Score the answer
Evaluate correctness and citation support, including abstention cases.
The practical take
Measure retrieval separately from answer quality so you know which component failed.
Sources & evidence
Capabilities above are summarized from these official documents. Workflow suggestions are editorial guidance, not a claim of measured product superiority.