To evaluate whether a retriever is returning relevant information independently of the generator’s performance, you need to focus on metrics and methods that isolate the retriever’s output from downstream processing. Start by measuring the retriever’s precision and recall against a labeled dataset. For example, if the query is “How to sort a list in Python,” manually verify if the top documents returned explain list sorting methods (e.g., sorted() vs. list.sort()). Use metrics like precision@k (how many of the top k results are relevant) and recall@k (how many relevant items were retrieved in the top k). These metrics don’t require the generator and directly reflect the retriever’s ability to surface correct information.
Another approach is to create a test set with predefined queries and known relevant documents. For instance, if a query asks about “SQL JOIN types,” the ideal retrieved documents should include explanations of INNER JOIN, LEFT JOIN, etc. By comparing the retriever’s output against this ground truth, you can calculate accuracy without involving the generator. Tools like Mean Reciprocal Rank (MRR) can also help quantify how high the first relevant document appears in the results. This method ensures the evaluation is deterministic and not influenced by the generator’s interpretation or potential errors.
Finally, conduct error analysis by examining cases where the generator produced incorrect or nonsensical answers. For example, if the generator answers a query about “Python decorators” incorrectly, check if the retriever provided documents about decorators or unrelated topics like variable scoping. If the retriever failed to fetch relevant content, the issue lies there. Additionally, run A/B tests by swapping retrievers (e.g., comparing BM25 vs. dense retrievers) while keeping the generator fixed. If performance changes significantly, it highlights the retriever’s impact. This layered evaluation isolates the retriever’s role and ensures improvements are targeted correctly.