In the industrial pre-training of large language models (LLMs), large-scale web content extraction faces a critical trade-off: rule-based methods sacrifice quality for efficiency, while model-based methods do the opposite. As a result, existing solutions deployed at scale often lead to significant data quality degradation and corpus loss. To address this challenge, which we faced directly at Ant Group, this paper introduces HiParse, a hierarchical adaptive web parsing framework deployed in production. The core of HiParse is to abstract this multi-dimensional challenge into a ''PESO'' (Precision, Efficiency, Scope, Overhead) evaluation model, quantifying parser capabilities and parsing requirements from a ''supply-demand'' perspective. The framework integrates a hybrid parser system and uses a lightweight decision engine to dynamically route different web pages to the optimal parsing strategy. Comprehensive experiments show that HiParse not only achieves superior parsing quality (9.03/10 score) but its direct output corpus, with only 30 billion tokens, also improved the average performance of a downstream LLM on 11 benchmarks by 7.6%. Furthermore, the framework has proven its high throughput, cost-effectiveness, and alignment with business parsing demands under real industrial loads. HiParse provides a high-fidelity, scalable, and cost-effective data solution for enterprise-level LLM pre-training. All the relevant models in the paper will be open source soon.