Letting the model write its own ripgrep commands to pull context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → matched graph-based retrieval that has to build and maintain a code index. It worked because the lines grep matched tended to sit physically near the spot being completed.
Plain text search breaks in two specific ways: high-frequency ambiguous keywords drag in noise, and fixed truncation boundaries cut retrieved code mid-structure so the model sees fragments.
GrepRAG fixes both with post-processing rather than a new index: re-rank matches by identifier weight, then dedupe along structural boundaries. That beat the best prior method by 7-15% relative on exact-match code completion.
The framing worth stealing: before paying for embeddingA list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.Full definition → or a code graph, measure how far index-free search already gets you, since the index is a permanent maintenance cost and grep is not.
Terms in this piece · Glossary
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Why it matters
Teams reach for embeddings and code graphs by default; this measures whether lexical search already gets there for less.
Key quotes
“Despite its simplicity, Naive GrepRAG achieves performance comparable to sophisticated graph-based baselines.”
“Further analysis shows that its effectiveness stems from retrieving lexically precise code fragments that are spatially closer to the completion site.”
“Extensive evaluation on CrossCodeEval and RepoEval-Updated demonstrates that GrepRAG consistently outperforms state-of-the-art (SOTA) methods, achieving 7.04-15.58 percent relative improvement in code exact match (EM) over the best baseline on CrossCodeEval.”
