DCI is a direct corpus interaction paradigm for agentic search. Instead of querying a fixed semantic retriever or retrieval API, the agent searches the raw corpus directly with terminal tools. This lets the agent freely compose search primitives and interact with the corpus as an open research environment. It also substantially simplifies the overall retrieval system.

DCI-Agent-Lite is the minimal open implementation of this paradigm, built on Pi with bash tools and lightweight context management for long-horizon deep research. With GPT-5.4-nano, it achieves an impressive 62.9% accuracy on BrowseComp-Plus, surpassing agentic search agents powered by GPT-5.2, Claude-Sonnet-4.6, Qwen3.5-122B, and GLM-4.7.

DCI-Agent-Lite outperforms top-performing baselines across 13 benchmarks spanning agentic search, knowledge-intensive QA, and IR-ranking tasks.

• Table 1 - Agentic Search results.

• Table 2 - Knowledge-intensive QA results.

• Table 3 - IR ranking results.

• 🔒 Your private deep-research assistant: Point DCI-Agent-Lite at a local corpus and start immediately. It searches, inspects, cross-checks, and answers from your own knowledge base without sending documents to a hosted retrieval service.
• ⚡ High-resolution, zero-index retrieval: No embeddings, vector databases, or offline index builds. The agent searches raw files directly with terminal commands like rg, find, and sed, so it can start immediately and maintain fine-grained control over the knowledge base.
• 🛠️ Minimal harness, long-horizon power: Built on Pi with only bash tools and lightweight context management, DCI-Agent-Lite is small enough to hack and strong enough for serious deep research runs.
• 🚀 Remarkable agentic-search performance: DCI-Agent-Lite with GPT-5.4-nano beats top baselines across 13 benchmarks, spanning BrowseComp-Plus, knowledge-intensive QA, and IR ranking.

• ⚙️ Setup
• ⚡ Quick Start
• 🚀 Running Experiments
• 🎯 Benchmark Evaluation
• 🏗️ Repository Layout
• 🙏 Acknowledgements
• 📚 Citation

Unix / macOS

bash setup.sh

# 1. Install uv + ripgrep, then sync Python deps
uv sync

# 2. Clone and build Pi
git clone https://github.com/jdf-prog/pi-mono.git pi-mono
cd pi-mono && git checkout codex/context-management-ablation && npm install && npm run build && cd ..

# 3. Configure API keys (copy template, edit .env, auto-loaded by setup.sh)
cp .env.template .env
# edit .env, then re-run setup.sh or source it manually

# 4. Download datasets (auto-downloaded by setup.sh, or run manually)
#    Corpus: https://huggingface.co/datasets/DCI-Agent/corpus
uv run python scripts/download_corpus.py

#    Benchmark datasets: https://huggingface.co/datasets/DCI-Agent/dci-bench
uv run python scripts/download_dci_bench.py

Copy the template to .env, then fill in the variables you need. To get DCI running, set at least one of OPENAI_API_KEY or ANTHROPIC_API_KEY:

cp .env.template .env

Common variables:

• OPENAI_API_KEY for OpenAI model runs and benchmark judging by default.
• ANTHROPIC_API_KEY for Anthropic model runs.

Prerequisites: Install dependencies and configure an OpenAI API key (see Setup).

The example below illustrates DCI-Agent-Lite in action: the deep research agent searches the corpus, inspects relevant documents, and produces evidence-grounded answers entirely within the given wikipedia corpus.

1. Open the DCI-Agent-Lite TUI:

# load keys from .env if not already in environment
set -a; source .env 2>/dev/null; set +a

uv run dci-agent-lite --terminal \
  --provider openai \
  --model gpt-5.4-nano \
  --cwd "corpus/wiki_corpus" \
  --extra-arg="--thinking high"

2. Run your first task. In the TUI, type:

Answer the following question using only wiki_dump.jsonl in the current directory. Do not use web search. Use rg instead of grep for fast searching. Question: In which street did the Great Fire of London originate?

3. (Optional) Run Programmatically from the CLI. Remove the --terminal flag and pass your task as the final argument:

set -a; source .env 2>/dev/null; set +a

uv run dci-agent-lite \
  --provider openai \
  --model gpt-5.4-nano \
  --cwd "corpus/wiki_corpus" \
  --extra-arg="--thinking high" \
  "Answer the following question using only wiki_dump.jsonl in the current directory. Do not use web search. Use rg instead of grep for fast searching. Question: In which street did the Great Fire of London originate?"

Programmatic runs save artifacts under outputs/runs/<timestamp>/. The final answer is in final.txt, the original question is in question.txt, and the full trajectory is in conversation_full.json. To choose a specific location, pass --output-dir path/to/run.

More runnable examples for OpenAI, Anthropic and vLLM are available in scripts/examples/ as dci_basic_*.sh. See the setup guide for vLLM configuration.

DCI-Agent-Lite includes a lightweight runtime context-management layer for long-horizon deep research runs.

It uses three simple strategies:

• Truncation shortens large tool results in each turn.
• Compaction keeps recent turns and replaces older tool results with placeholders.
• Summarization summarizes older history when the context gets crowded.

Context management illustration

The runtime levels move from no context management to more aggressive compression:

Pass a level through Pi with --extra-arg:

set -a; source .env 2>/dev/null; set +a

uv run dci-agent-lite \
  --provider openai \
  --model gpt-5.4-nano \
  --cwd "corpus/wiki_corpus" \
  --extra-arg="--thinking high" \
  --extra-arg="--context-management-level level4" \
  "Answer the following question using only wiki_dump.jsonl in the current directory. Do not use web search. Use rg instead of grep for fast searching. Question: In which street did the Great Fire of London originate?"

We benchmark DCI-Agent-Lite on the following benchmark suites using OpenAI gpt-5.4-nano with --thinking high, context management set to level3, and a maximum turn budget of 300.

bash scripts/bcplus_eval/run_bcplus_eval_openai.sh

bash scripts/qa/run_hotpotqa_dev_sample50.sh
bash scripts/qa/run_musique_dev_sample50.sh
bash scripts/qa/run_nq_test_sample50.sh
bash scripts/qa/run_triviaqa_test_sample50.sh
bash scripts/qa/run_2wikimultihopqa_dev_sample50.sh
bash scripts/qa/run_bamboogle_test_sample50.sh

# BRIGHT
bash scripts/bright/run_bio.sh
bash scripts/bright/run_earth_science.sh
bash scripts/bright/run_economics.sh
bash scripts/bright/run_robotics.sh

Zhuofeng Li

Dongfu Jiang

Haoxiang Zhang

Cong Wei

Pan Lu

Ping Nie

Yejin Choi

James Zou

Jiawei Han

Wenhu Chen

Jimmy Lin

Yu Zhang

@article{li2026beyond,
  title={Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction},
  author={Li, Zhuofeng and Zhang, Haoxiang and Wei, Cong and Lu, Pan and Nie, Ping and Lu, Yi and Bai, Yuyang and Feng, Shangbin and Zhu, Hangxiao and Zhong, Ming and Zhang, Yuyu and Xie, Jianwen and Choi, Yejin and Zou, James and Han, Jiawei and Chen, Wenhu and Lin, Jimmy and Jiang, Dongfu and Zhang, Yu},
  journal={arXiv preprint arXiv:2605.05242},
  year={2026}
}

↑ Back to Top ↑