Can direct corpus search beat embedding-based retrieval?

The standard search-agent stack assumes a retriever: a query goes in, a precomputed index returns a ranked list, and the agent reasons over the results. GrepSeek inverts the substrate — the corpus itself is the environment, and the agent finds evidence by issuing executable shell commands (grep and friends) over raw text. The motivation is a specific failure of dense retrieval: embedding-based models semantically conflate distinct entities, so multi-hop queries that hinge on exact symbolic patterns or strict entity-level constraints get the wrong documents. Direct corpus interaction (DCI) recovers the lexical precision that dense vectors blur, and as a side effect eliminates the expensive offline indexing stage and reduces runtime memory via sharded-parallel execution.

The interesting part is the training, because RL directly on a large corpus is unstable — the action space of shell commands over raw text is enormous and rewards are sparse. GrepSeek uses a two-stage pipeline: an answer-aware Tutor and an answer-blind Planner generate verified, causally grounded cold-start trajectories for SFT, then GRPO refines task-oriented search behavior through direct interaction. This is the same architectural move as Can externalized bookkeeping let smaller search agents beat larger ones? applied to the retrieval primitive: scaffold the unstable parts so RL only has to learn the genuinely strategic search behavior.

It would be wrong to read this as "grep beats embeddings." The win is regime-specific — symbolic, entity-constrained, lexically exact needs. Dense retrieval still dominates paraphrase and conceptual matching, where the query and the answer share meaning but no surface tokens; an exact-match agent is blind exactly there. So GrepSeek is best understood as restoring a lexical-precision tool that dense-only pipelines discarded, a complementary perspective rather than a replacement. It also complicates the convenience story behind Why do search agents fail users despite strong benchmark scores?: interpretable, inspectable retrieval programs are auditable in a way a dense top-k is not, but they push the burden of expressing the need into precise command construction the model must learn to author.