Add text-retrieval task category and improve documentation

README.md

@@ -1,17 +1,75 @@
 ---
+task_categories:
+- text-retrieval
 tags:
 - agent
 ---
+
 This dataset hosts the [AgentIR-4B](https://huggingface.co/Tevatron/AgentIR-4B) indexes.
-- Paper: https://arxiv.org/abs/2603.04384
-- Code: https://github.com/texttron/AgentIR/tree/main
-- Model: https://huggingface.co/Tevatron/AgentIR-4B
-- Project Page: https://texttron.github.io/AgentIR/
 
-For usage details of this index, please see https://github.com/wu-ming233/AgentIR-dev/tree/main/evaluation.
+- **Paper:** [AgentIR: Reasoning-Aware Retrieval for Deep Research Agents](https://huggingface.co/papers/2603.04384)
+- **Code:** [https://github.com/texttron/AgentIR](https://github.com/texttron/AgentIR)
+- **Model:** [Tevatron/AgentIR-4B](https://huggingface.co/Tevatron/AgentIR-4B)
+- **Project Page:** [https://texttron.github.io/AgentIR/](https://texttron.github.io/AgentIR/)
 
-## Citation
+For usage details of this index, please see [https://github.com/wu-ming233/AgentIR-dev/tree/main/evaluation](https://github.com/wu-ming233/AgentIR-dev/tree/main/evaluation).
+
+## Quick Usage
+
+Below is the example code from the official repository to embed queries (including reasoning) and documents using the AgentIR-4B model:
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+MODEL = "Tevatron/AgentIR-4B"
+PREFIX = "Instruct: Given a user's reasoning followed by a web search query, retrieve relevant passages that answer the query while incorporating the user's reasoning
+Query:"
+QUERY = """Reasoning: Search results show some relevant info about music and Grammy. We need a composer who won a Grammy, could be from Sweden/Finland/Austria (joined 1995)? The person is known for a certain creation that is a subgenre known for euphoric finale. Which subgenre has a euphoric finale? "Progressive house"? There's a structure: Build-up, breakdown, climax, drop, euphoria. They started creating this piece in a small studio's backroom.
+
+Query: "backroom" "studio" "early 2010s" "euphoric"
+"""
+DOCS = [
+    "35+ Studios With Upcoming Games to Watch: Turtle Rock Studios
+
+Making its name on the classic Left 4 Dead series of games, Turtle Rock Studios is working on an all-new co-op game called Back 4 Blood that sees you fighting through a zombie apocalypse. Sound familiar? Announced in early 2019 and being published",
+    "name: Otto Knows
+image_upright: 1.25
+birth_name: Otto Jettman
+birth_date: 6 05 1989
+birth_place: Stockholm, Sweden
+genre: Electro house, house, progressive house
+occupation: DJ, music producer, remixer
+
+Otto Jettman (born 6 May 1989), better known by his stage name Otto Knows is a Swedish DJ, producer and remixer who has had a number of hits in Sweden, Belgium and the Netherlands"
+]
+
+def embed(texts, model, tokenizer, device, is_query=False):
+    batch = tokenizer(
+        [PREFIX + t if is_query else t for t in texts],
+        padding=True,
+        truncation=True,
+        max_length=8192,
+        return_tensors="pt",
+    )
+    batch = {k: v.to(device) for k, v in batch.items()}
+    with torch.no_grad():
+        hidden = model(**batch, return_dict=True).last_hidden_state
+        reps = hidden[:, -1]
+        return torch.nn.functional.normalize(reps, p=2, dim=-1).cpu()
+
+model = AutoModel.from_pretrained(MODEL, torch_dtype=torch.float16, device_map="auto")
+device = model.device
+tokenizer = AutoTokenizer.from_pretrained(MODEL, padding_side="left")
+
+q = embed([QUERY], model, tokenizer, device, is_query=True)[0]
+docs = embed(DOCS, model, tokenizer, device)
+for doc, vec in zip(DOCS, docs):
+    print(f"{torch.dot(q, vec).item():.6f}  {doc}")
 ```
+
+## Citation
+```bibtex
 @article{chen2026AgentIR,
       title={AgentIR: Reasoning-Aware Retrieval for Deep Research Agents}, 
       author={Zijian Chen and Xueguang Ma and Shengyao Zhuang and Jimmy Lin and Akari Asai and Victor Zhong},