---
pretty_name: AgentWebBench Corpus
license: mit
language:
  - en
task_categories:
  - text-retrieval
size_categories:
  - 10M<n<100M
tags:
  - information-retrieval
  - dense-retrieval
  - faiss
  - embeddings
  - clueweb22
  - agents
  - benchmark
---

# AgentWebBench Corpus

Pre-built **dense-retrieval corpus** for [AgentWebBench](https://arxiv.org/abs/2604.10938) [ICML 2026], a benchmark for Multi-Agent Coordination in Agentic Web over a realistic 100-website slice of
[ClueWeb22](https://lemurproject.org/clueweb22/) (~18.4M documents).

This repository holds the **embeddings and FAISS indices** the benchmark loads at run time, including per-website indices, a global index, and website-level vectors. It does **not** contain ClueWeb22 text (see [Raw documents](#raw-documents-clueweb22-b)).
- **Websites:** 100
- **Documents:** ~18.4M 
- **Embedding dim:** 1024


⚠️ **Derived from ClueWeb22.** These vectors and ID maps are derived from ClueWeb22 documents. Use is subject to the [ClueWeb22 license](https://lemurproject.org/clueweb22/).

## 1. Contents

```
faiss_indices/
├── <website>.faiss            # per-website FAISS index (IndexFlatIP over doc vectors)
└── <website>_doc_ids.npy      # row i of the index → that website's ClueWeb22 doc_id
website_embeddings.pkl         # {website → 1024-d vector}; used to rank/select websites
global_doc_index.faiss         # one FAISS index over ALL ~18.4M documents
global_doc_ids.npy             # row i of the global index → ClueWeb22 doc_id
```

**How `.faiss` and `_doc_ids.npy` pair up:** a FAISS search returns integer row positions, not document IDs. The `.faiss` stores the vectors; the `_doc_ids.npy` is the row→`doc_id` lookup. They are strictly aligned by position and must be used together.


## 2. Usage

### 2.1 With the AgentWebBench code (recommended)

```bash
git clone https://github.com/cxcscmu/AutoGEO/AgentWebBench
cd AgentWebBench
python -m awbench.download --output-dir ./AgentWebBench-corpus
# then set WEBSITE_DATA_ROOT in awbench/config.py to ./AgentWebBench-corpus
```

### 2.2 Directly with `huggingface_hub`

```python
from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="cx-cmu/AgentWebBench-corpus",
    repo_type="dataset",
    local_dir="./AgentWebBench-corpus",
)
```

### 2.3 Load an index

```python
import faiss, numpy as np

index = faiss.read_index("faiss_indices/community.spiceworks.com.faiss")
doc_ids = np.load("faiss_indices/community.spiceworks.com_doc_ids.npy", allow_pickle=True)

# encode your query with MiniCPM-Embedding-Light (1024-d, normalized), then:
scores, rows = index.search(query_vec.reshape(1, -1).astype("float32"), 10)
hits = [str(doc_ids[i]) for i in rows[0]]   # ClueWeb22 doc_ids
```


This corpus stores only vectors and `doc_id`s. To read the actual page text for a retrieved `doc_id`, obtain [ClueWeb22 category B](https://lemurproject.org/clueweb22/index.php) (license required) and point `CLUEWEB_ROOT_PATH` at it.

## 3. How it was built

We follow [tevatron](https://github.com/texttron/tevatron) to build embeddings and FAISS indices. Specifically,
- **Encoder:** [`openbmb/MiniCPM-Embedding-Light`](https://huggingface.co/openbmb/MiniCPM-Embedding-Light)
- **Dimension:** 1024 
- **Index:** FAISS `IndexFlatIP`

## 4. Citation

```bibtex
@article{zhong2026agentwebbench,
  title={AgentWebBench: Benchmarking Multi-Agent Coordination in Agentic Web},
  author={Zhong, Shanshan and Shen, Kate and Xiong, Chenyan},
  journal={arXiv preprint arXiv:2604.10938},
  year={2026}
}
```