GrepSeek: Training Search Agents for Direct Corpus Interaction

GrepSeek trains compact search agents to query raw text corpora directly with shell commands instead of relying on traditional indexed retrievers.

It uses a two-stage pipeline: a verified cold-start dataset built by an answer-aware Tutor and answer-blind Planner, then GRPO fine-tuning.

A sharded-parallel execution engine speeds up shell-based retrieval while preserving exact results.

Across seven open-domain QA benchmarks, GrepSeek achieves the best overall token-level F1 and Exact Match.