README.md

metadata

library_name: transformers
pipeline_tag: text-generation

PyRAG-7b

PyRAG is a framework that reformulates multi-hop Retrieval-Augmented Generation (RAG) as program synthesis and execution. Instead of representing reasoning as free-form natural language, PyRAG decomposes a question into atomic sub-queries, synthesizes an executable Python program over tool primitives — retrieve(query) and answer(query, docs) — and runs the program step-by-step in a Python interpreter.

This repository contains the 7B parameter model checkpoint presented in the paper Retrieval is Cheap, Show Me the Code: Executable Multi-Hop Reasoning for Retrieval-Augmented Generation.

Resources

• Paper: https://arxiv.org/abs/2605.12975
• GitHub Repository: https://github.com/GasolSun36/PyRAG
• Project Page: https://gasolsun36.github.io/PyRAG/

Sample Usage

You can use PyRAG as a library by following the installation instructions in the official repository. Below is the programmatic usage example:

from pyrag import (
    HttpRetrievalAgent,
    OpenAILLM,
    RAGProgramRunner,
    env_enable_thinking,
)

instruct_llm = OpenAILLM(
    model="Qwen/Qwen2.5-7B-Instruct",
    base_url="http://127.0.0.1:8337/v1",
    enable_thinking=env_enable_thinking(),
)
plan_llm = OpenAILLM(
    model="Qwen/Qwen2.5-Coder-7B-Instruct",
    base_url="http://127.0.0.1:8336/v1",
    enable_thinking=env_enable_thinking(),
)
retrieval_agent = HttpRetrievalAgent(host="127.0.0.1", port=8008)

runner = RAGProgramRunner(
    llm=instruct_llm,
    plan_llm=plan_llm,
    retrieval_agent=retrieval_agent,
)

result = runner.run(
    "How old was Virginia Bruce when she starred in Let Freedom Ring?",
    topk=5,
)

print(result["final_answer"])         # → 29
print(result["sub_queries"])          # decomposed atomic queries
print(result["generated_code"])       # the synthesized Python program
print(result["execution_log"])        # full step-by-step trace
print(result["retried_with_topk10"])  # whether adaptive retrieval triggered

Citation

@misc{sun2026retrievalcheapcodeexecutable,
      title={Retrieval is Cheap, Show Me the Code: Executable Multi-Hop Reasoning for Retrieval-Augmented Generation}, 
      author={Jiashuo Sun and Jimeng Shi and Yixuan Xie and Saizhuo Wang and Jash Rajesh Parekh and Pengcheng Jiang and Zhiyi Shi and Jiajun Fan and Qinglong Zheng and Peiran Li and Shaowen Wang and Ge Liu and Jiawei Han},
      year={2026},
      eprint={2605.12975},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.12975}, 
}