An acceptable range for retriever recall in a RAG system aiming to answer questions correctly "most of the time" typically falls between 70% and 90%, depending on the domain and use case. Recall measures the retriever’s ability to find all relevant documents in a corpus. Higher recall reduces the risk of the generator missing critical information, but it can also introduce irrelevant content, which might degrade answer quality if the generator struggles to filter noise. For example, in a general-purpose QA system targeting 80-90% end-to-end accuracy, a retriever recall of 75-85% might suffice if the generator is robust enough to handle some missing context. However, this range isn’t universal and must align with the generator’s capabilities and the application’s tolerance for errors.
The required recall varies significantly by domain. In high-stakes fields like healthcare or legal research, missing even a single critical document could lead to harmful outcomes. Here, recall should be prioritized (e.g., 85-95%), often requiring domain-specific tuning. For instance, a medical RAG system might use specialized embeddings trained on clinical texts to improve retrieval of rare conditions. In contrast, customer support chatbots might tolerate lower recall (65-80%) because questions are often repetitive, and the generator can infer answers from partial data. Similarly, in enterprise search, where documents are structured and queries are predictable, moderate recall (70-85%) might be acceptable if metadata or keyword filters reduce the search space.
Other factors influencing the target recall include corpus size and document density. A system querying a small, curated knowledge base (e.g., a company’s internal docs) might achieve 90% recall with simple keyword matching. However, a web-scale corpus with diverse content might require hybrid retrieval (e.g., combining dense vectors with sparse methods) to hit 70% recall. Additionally, recall can be traded off against latency: real-time applications might limit the number of retrieved documents, lowering recall but improving speed. Balancing these trade-offs requires iterative testing, using metrics like answer accuracy on domain-specific benchmarks to validate the retriever’s performance.