Can Agents Search Without a Vector Database? A Simple Grep Is Enough

As large language models shift from answering questions to performing self‑search, verification, and evidence synthesis, traditional retrieval pipelines—chunking documents, building vector indexes, and returning a top‑k list—often miss critical weak clues needed for complex tasks.

The authors propose Direct Corpus Interaction (DCI), a new search paradigm that eliminates embedding models and vector indexes. Agents interact directly with the raw corpus using tools such as grep, rg, find, glob patterns, and lightweight scripts to locate, verify, and refine hypotheses.

Experiments on the BrowseComp‑Plus benchmark with Claude Sonnet 4.6 show that replacing the Qwen3‑Embedding‑8B retriever with DCI raises accuracy from 69.0 % to 80.0 % while reducing cost from $1440 to $1016. On multi‑hop QA datasets (NQ, TriviaQA, Bamboogle, HotpotQA, 2WikiMultiHopQA, MuSiQue), DCI‑Agent‑CC attains an average 83.0 % accuracy, a 30.7 % gain over the strongest baseline ASearcher‑Local‑14B. On IR benchmarks (BRIGHT, BEIR) it achieves 68.5 % average NDCG@10, surpassing ReasonRank‑32B by 21.5 %.

Analysis of agent trajectories reveals that DCI‑Agent‑Lite covers fewer relevant documents than the embedding‑based retriever but localizes evidence far more precisely, demonstrating that finer‑grained “retrieval interface resolution”—exposing file names, line numbers, match snippets, and context—drives the performance boost.

The paper implements two agents: DCI‑Agent‑Lite, a lightweight CLI agent using bash and basic tools, and DCI‑Agent‑CC, built on Claude Code with richer prompting and tool orchestration. To handle long‑running searches, the agents employ three context‑management mechanisms: truncation of oversized tool outputs, compression of historic rounds by replacing raw results with placeholders, and summarization agents that rewrite long histories into structured notes.

However, DCI is not a universal solution. When corpus size grows, the number of file scans, tool invocations, latency, and cost increase substantially, making DCI less suitable for massive open‑world retrieval. It excels on local, heterogeneous, rapidly changing corpora such as code repositories, research archives, or internal enterprise documents that require multi‑round exploration and strict evidence verification.

Future systems may combine coarse‑recall traditional retrievers with DCI’s high‑resolution verification, using the former to narrow the search space and the latter to perform precise, evidence‑driven grounding.