Open-Dataflow/dataflow-demo-AgenticRAG

Agentic RAG数据合成管道是一个端到端的框架，旨在支持基于强化学习的代理RAG训练，并从给定的文本内容生成高质量的问答对。该管道包括两个主要步骤：1. 原子任务生成，用于从原始文档生成问题、参考答案、精炼参考答案、可选的可验证答案以及LLM对问题的答案。2. 任务质量评估，用于评估给定原始文档的精炼参考答案和LLM答案之间的F1分数，以确保每个构建的问题在配对正确的文档检索时都能获得适当的奖励，从而保持强化学习的训练质量。

The Agentic RAG Data Synthesis Pipeline is an end-to-end framework designed to support RL-based agentic RAG training and generate high-quality question-answer pairs from provided text contents. The pipeline includes two main steps: 1. Atomic Task Generation, which generates questions, reference answers, refined reference answers, optional verifiable answers, and the LLMs answer to the question from the original document. 2. Task Quality Evaluation, which evaluates the F1 score between the refined reference answer and the LLMs answer to the question given the original document, to ensure that each constructed question receives an appropriate reward when paired with correct document retrieval, maintaining the training quality of reinforcement learning.