Semantics Delivery Network: Rethinking Web Retrieval Infrastructure for LLM Agents

Overview

Semantics Delivery Network (SemDN) proposes an origin-authorized edge layer that retrieves and delivers web content as task-relevant chunks for LLM agents. It addresses the gap between URL-based search and caching infrastructure and the short, useful passages agents consume across multi-step tasks.

Method

A client sends a semantic query to a nearby edge, which searches a hierarchical corpus and returns selected chunks. SemDN shares origin-authorized content acquisition and normalization, supports tenant-specific ranking, and caches normalized blocks, chunks, and index state. A coverage controller estimates when the corpus may be incomplete or stale and can trigger scoped discovery or refresh.

Evaluation

In preliminary measurements on 100 live-web queries, agents consumed a median 7.1 KB of selected chunks while lightweight HTML retrieval processed 761 KB to obtain them (107×); full browser renders processed 18.1 MB. In trace-driven cache replays of Search-R1 over wiki-18, a 1.34 MB chunk cache served 73.8% of chunk requests and sent 13–41× fewer bytes to origin than a URL-object cache at equal capacity. On a fixed wiki-18 corpus across five QA sets, chunk-first retrieval with the same Qwen2.5-7B reader improved average F1 from 0.334 to 0.357 over page-first retrieval while delivering about 470× fewer bytes.

Citation

@misc{hua2026semantics,
  title={Semantics Delivery Network: Rethinking Web Retrieval Infrastructure for LLM Agents},
  author={Peichun Hua and Yunming Xiao},
  year={2026},
  eprint={2609.22486},
  archivePrefix={arXiv},
  primaryClass={cs.NI},
  url={https://arxiv.org/abs/2609.22486}
}