Introduction
This is the Agentic-R trained in our paper: Agentic-R: Learning to Retrieve for Agentic Search (📝arXiv). Please refer our 🧩github repository for the detailed usage of our Agentic-R.
Usage
Our Agentic-R query encoder is designed for agentic search scenarios.
For queries, the input format is:
query: <original_question> [SEP] <agent_query>.
Passages use the standard passage: prefix following E5.
Below is an example of how to compute embeddings using sentence_transformers:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("liuwenhan/Agentic-R_e5")
input_texts = [
# Query encoder input:
# original_question [SEP] current_query
"query: Who wrote The Old Man and the Sea? [SEP] Old Man and the Sea",
# Passages
"passage: The Old Man and the Sea is a short novel written by the American author Ernest Hemingway in 1951.",
"passage: Ernest Hemingway was an American novelist, short-story writer, and journalist, born in 1899."
]
embeddings = model.encode(
input_texts,
normalize_embeddings=True
)
Notes:
original_question refers to the user’s initial question.
agent_query refers to the intermediate query generated during the agent’s reasoning process.
Always include [SEP] to separate the two parts of the query.
We recommend setting normalize_embeddings=True for cosine similarity–based retrieval.