A robot that browses files to find answers works fine on a small library, but how big before plain keyword search wins?
How many books does it take before raw navigation collapses completely?
This explores at what corpus size an AI agent that finds things by browsing files and reading them one at a time stops being competitive with a plain keyword index, with 'books' as the unit of corpus size.
This explores at what corpus size an agent that finds things by browsing files and reading them one at a time stops keeping up with a plain keyword index, using books as the yardstick. The corpus gives a crossover point in tokens, not in books: Does lexical search outperform agentic navigation as corpus size grows? finds that BM25, the classic keyword-ranking method, overtakes a file-system agent at around 10 million corpus tokens. A typical book runs about 100,000 tokens, which is my own rough conversion and not a figure from the corpus. That puts the crossover at roughly a hundred books, or a modest shelf.
The corpus doesn't describe a complete collapse, though. The finding is a handoff. Agent navigation leads on small corpora, becomes inefficient as the search space grows, and past the crossover BM25 keeps the accuracy and cost advantage. The reason is structural. An agent walks the collection sequentially, opening one thing at a time, while a global index shows candidates from the whole corpus at once. Adding books makes the walk longer, but it doesn't change how long a lookup in the index takes to set up.
The lesson is not that reading loses to searching. Can long-context models resolve retriever-reader imbalance? shows the best RAG designs now do coarse ranking first and then deep reading, with 4K-token chunks handed to a long-context reader in place of 100-word snippets. Cheap global narrowing decides where the reader looks, and the reader still does the careful part.
Two neighbouring ideas make the same point from other angles. Can building a document map first improve retrieval over long texts? builds a map of a long document before retrieving from it, so scattered evidence can be found by its role in the whole. That paper works within long documents, not shelves of books, but it shows the same fix of giving the searcher a map first. Why does vanilla RAG produce shallow and redundant results? adds a warning: a single pass of retrieval keeps mining one semantic neighbourhood, so iterating over different queries matters as much as raw scale.
The corpus doesn't say how the crossover moves with book length, model strength or task type, or where agent navigation stops being usable at all. It supports one number, about 10 million tokens, and one direction: past a few dozen to a hundred books, browsing without an index is the wrong place to start.
Sources 4 notes
Lexical search indexed globally outperforms sequential agentic exploration at scale because it exposes relevant candidates across the entire corpus at once. Agent-based search leads only on small corpora but becomes inefficient as search space grows, while BM25 maintains accuracy and cost advantage at deployment scale.
LongRAG shows that 4K-token units and long-context readers outperform 100-word retrieval on standard benchmarks. The optimal RAG design shifts from precise retrieval to coarse ranking plus deep reading as context windows expanded.
MiA-RAG inverts standard RAG by summarizing documents first, then conditioning retrieval on that global view. This approach recovers discourse structure that bag-of-chunks retrieval destroys, making scattered evidence findable by their document role rather than surface similarity alone.
Vanilla RAG fails not at retrieval quality but retrieval diversity—it exploits one semantic neighborhood repeatedly. Iterative expansion-reflection cycles, which regenerate queries based on cognitive reorganization, mirror human reflective practice and raise knowledge density by traversing multiple knowledge neighborhoods.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LongRAG: A Dual-Perspective Retrieval-Augmented Generation Paradigm for Long-Context Question Answering
- LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs
- A New Role for Relevance: Guiding Corpus Interaction in Agentic Search
- Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs
- A Hybrid RAG System with Comprehensive Enhancement on Complex Reasoning
- You Don't Need Pre-built Graphs for RAG: Retrieval Augmented Generation with Adaptive Reasoning Structures
- CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning
- BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms