Long-Context RAG

The Question

If a model’s context window grows to 32K, 128K or a million tokens, is retrieval still necessary — or can the whole corpus simply be pasted in?

The Empirical Answer: Both

NVIDIA Research (covered by Ravi Theja for LlamaIndex, October 2023) evaluated Nemo GPT-43B and LLaMA2-70B at extended 16K and 32K contexts across seven datasets — QMSum, Qasper, NarrativeQA, QuALITY, MuSiQue, HotpotQA and MultiFieldQA-en. The findings:

  • Retrieval significantly improves both short-context (4K) and long-context (16K/32K) models. Long context does not make retrieval redundant.
  • LLaMA2-70B-32K with retrieval surpassed GPT-3.5-turbo-16K and matched davinci-003.
  • A 4K model with retrieval performs comparably to a much longer model, at faster inference.

The Lost-in-the-Middle Constraint

Retrieving more is not monotonically better. Top 5–10 chunks was optimal; retrieving 20 chunks degraded performance, attributed to the “lost in the middle” phenomenon — material placed in the middle of a long context is attended to less reliably than material at either end.

This is the same effect the Needle in a Haystack Test measures directly, and it is why long context does not dissolve the ranking problem: position within the context window is itself a ranking decision.

Practical Consequence

Long context changes the budget, not the task. You still need to decide what goes in and in what order — and Prompt Compression becomes attractive precisely because filling a large window is expensive.

Articles