Ask a grounded question over the indexed LangChain documentation. Answers include retrieved passages, ranks, and source citations.
Offline evaluation
Retrieval quality
Measured against a labeled LangSmith documentation question set.
Loading snapshot…
Recall curve
Hit rate by cutoff
Dataset slices
Performance by difficulty
No matching questions.
Retrieval Hit@5 is high because the 100 questions were written from labeled source quotes. Faithfulness is unigram overlap of generated test-split answers with top-5 chunks — paraphrase-tolerant, not LLM-as-judge. Stricter bigram precision on the same answers is lower, because wording does not have to copy the source.