Generate rather than Retrieve: Large Language Models are Strong Context Generators

Knowledge-intensive tasks, such as open-domain question answering (QA), require access to a large amount of world or domain knowledge. A common approach for knowledge-intensive tasks is to employ a retrieve-then-read pipeline that first retrieves a handful of relevant contextual documents from an external corpus such as Wikipedia and then predicts an answer conditioned on the retrieved documents. In this paper, we present a novel perspective for solving knowledge-intensive tasks by replacing document retrievers with large language model generators. We call our method generate-then-read (GenRead), which first prompts a large language model to generate contextutal documents based on a given question, and then reads the generated documents to produce the final answer. Furthermore, we propose a novel clustering-based prompting method that selects distinct prompts, resulting in the generated documents that cover different perspectives, leading to better recall over acceptable answers. We conduct extensive experiments on three different knowledge-intensive tasks, including open-domain QA, fact checking, and dialogue system. Notably, GenRead achieves 71.6 and 54.4 exact match scores on TriviaQA and WebQ, significantly outperforming the state-of-the-art retrieve-then-read pipeline DPR-FiD by +4.0 and +3.9, without retrieving any documents from any external knowledge source. Lastly, we demonstrate the model performance can be further improved by combining retrieval and generation. Our code and generated documents can be found at https://github.com/wyu97/GenRead.

1. Generate-then-Read

Instead of the traditional “retrieve-then-read” pipeline, they propose a “generate-then-read” pipeline for RAG. This means generating a hypothetical document that may contain the answer, and using that to answer the question.

Instead of generating just a single document, they used a variety of techniques to increase diversity of the generated documents, such as diverse human-written prompts and sampling random few-shot examples of question-document pairs to seed it. There’s nothing really novel about the technique, but the surprising thing is that this actually works for popular Q&A datasets:

Most Glaring Deficiency

Some evals on proprietary/unseen datasets would be interesting. We all know it probably wouldn’t work, but it would be nice to confirm this, or otherwise be pleasantly surprised.

Conclusions for Future Work

Kind of a silly idea prima facie, but it’s quite cool that it works in some domains. Kudos to the authors for putting in the effort to try this out.