Retrieval meets Long Context Large Language Models

Extending the context window of large language models (LLMs) is getting popular recently, while the solution of augmenting LLMs with retrieval has existed for years. The natural questions are: i) Retrieval-augmentation versus long context window, which one is better for downstream tasks? ii) Can both methods be combined to get the best of both worlds? In this work, we answer these questions by studying both solutions using two state-of-the-art pretrained LLMs, i.e., a proprietary 43B GPT and LLaMA2-70B. Perhaps surprisingly, we find that LLM with 4K context window using simple retrieval-augmentation at generation can achieve comparable performance to finetuned LLM with 16K context window via positional interpolation on long context tasks, while taking much less computation. More importantly, we demonstrate that retrieval can significantly improve the performance of LLMs regardless of their extended context window sizes. Our best model, retrieval-augmented LLaMA2-70B with 32K context window, outperforms GPT-3.5-turbo-16k and Davinci003 in terms of average score on seven long context tasks including question answering and query-based summarization. It also outperforms its non-retrieval LLaMA2-70B-32k baseline by a margin, while being much faster at generation. Our study provides general insights on the choice of retrieval-augmentation versus long context extension of LLM for practitioners.

・4k+RAGなら16kとほぼ同じ精度でより高速に推論できる

・RAG有りLLaMA2-70B-32kはGPT-3.5-turbo-16kとDavinci003を凌駕しベスト性能

https://pbs.twimg.com/media/F70uxoWXMAELZPN?format=jpg&name=medium#.png

@nvidia

compares retrieval augmentation generation (RAG) with increasing context window for large language models (LLMs) on long context question answering and summarization tasks.

https://pbs.twimg.com/media/F70uxoWXMAELZPN?format=jpg&name=medium#.png

Extended context window to 16k for GPT-43B and 16K/32k for Llama using positional interpolation using the PILE dataset

Instruction-tuned small and big context models on QA and summarization

Evaluated on 7 datasets

Compare the performance of LLM + Retrieval for 4k & 16k/32k against non-retrieval with full context. Also, compare to GPT-3.5-turbo

The chunk size for embedding was 300 words

The best performance of RAG was for 5-10 chunks

Simple RAG + 4k LLM can match long context LLM.

RAG + 32k LLM performs better than providing the full context

Positional interpolation is a simple method to extend context length by ~4-8x