An execution engine runs commands in parallel across sharded threads.
Andrew Drozdov#1066
1/n Do search agents always need index-based retrievers to work efficiently & effectively?🤔Maybe not, if you TEACH them to interact with the corpus by shell!🤯
GrepSeek, a paradigm for training fast & practical Direct Corpus Interaction Search Agents!🚀 https://huggingface.co/papers/2605.29307
Many users praise GrepSeek for exceeding expectations with higher F1 scores on QA tasks and for offering a competitive practical approach to real-world multi-hop reasoning via shell-based agents.
No Digg Deeper questions have been answered for this story yet.
the information retrieval community has long known that structured queries provide more precise and surgical interaction with a corpus compared to keywords. together with folks at umass, led by @SalemiAlireza7, we show the effective and efficient use of the same tools by agents.
1/n Do search agents always need index-based retrievers to work efficiently & effectively?🤔Maybe not, if you TEACH them to interact with the corpus by shell!🤯
GrepSeek, a paradigm for training fast & practical Direct Corpus Interaction Search Agents!🚀 https://huggingface.co/papers/2605.29307
There’s a major shift underway in information retrieval: search agents with broader capabilities and access to more tools than any systems before them.
Follow @HamedZamani and collaborators to stay close to the bleeding edge.
1/n Do search agents always need index-based retrievers to work efficiently & effectively?🤔Maybe not, if you TEACH them to interact with the corpus by shell!🤯
GrepSeek, a paradigm for training fast & practical Direct Corpus Interaction Search Agents!🚀 https://huggingface.co/papers/2605.29307
6/n Results? GrepSeek performs better what we even expected! 🏆 A higher avg F1 compared to retrieval-based agent baseline on 7 open-domain QA datasets! 📈
4/n Training was no joke. Naive RL leads to unstable behavior and training collapse 😬
A cold-start dataset built with an answer-aware Tutor and an answer-blind Planner in backward and forward evidence collection paths , followed by GRPO refinement! 🧠🔥
5/n "But grep is too slow for 21M docs and not practical!" 🐢
Not anymore! We built a semantics-preserving sharded-parallel execution engine that accelerates shell-based retrieval by up to 7.6x! 🏎️💨
Average search latency drops from 5.39 seconds to just 0.71 seconds! ⏱️
2/n Instead of relying on pre-computed document indices, GrepSeek agents treat the corpus itself as the search environment!ℹ️
GrepSeek finds evidence by issuing executable shell commands! 💻
3/n We trained a compact (9B) LLM to search, filter, and compose evidence from massive unstructured text (14GB) using tools like rg and head. 🐧
It performs "surgical" information retrieval at any granularity without being restricted by pre-determined text chunks! ✂️
7/n GrepSeek is a competitive and practical approach for (at least) multi-hop reasoning in the real world. 🌍
Codebase, training data, and model checkpoints are officially live! 🛠️
Code: https://github.com/alirezasalemi7/grepseek Checkpoints: https://huggingface.co/collections/alireza7/grepseek Paper: https://arxiv.org/abs/2605.29307
8/n Shoutout to @SalemiAlireza7 for leading the project! You should follow him, if you haven’t already!