LARAG: Link-Aware Retrieval Strategy for RAG Systems in Hyperlinked Technical Documentation
LARAG introduces a link-aware retrieval strategy that improves RAG systems by leveraging hyperlink structures already present in technical documentation, rather than treating documents as flat text collections. The approach achieves better answer quality with fewer computational resources, demonstrating that implicit graph-like retrieval through existing metadata can enhance AI system performance.
LARAG addresses a fundamental inefficiency in how retrieval-augmented generation systems process structured digital content. Most RAG implementations flatten hyperlinked documents into isolated passages, discarding the navigational relationships that authors embedded through links. This research demonstrates that preserving these author-defined connections significantly improves both retrieval accuracy and computational efficiency, validated through benchmarking on technical documentation with expert-designed queries.
The broader context reflects growing recognition that LLMs require better grounding mechanisms beyond simple embedding similarity. As enterprises deploy RAG for internal knowledge bases, product documentation, and technical support, the challenge of handling structured corpora effectively becomes critical. Traditional embedding-based retrieval treats all passages equally, missing the semantic importance hierarchy that link structures communicate. LARAG's lightweight approach requires no separate graph construction or inference overhead, making it immediately applicable to existing HTML documentation without architectural changes.
For developers and enterprises, this work has practical implications. Organizations using RAG for documentation retrieval can improve answer quality while reducing computational costs by leveraging existing HTML structures. The approach scales efficiently since it exploits metadata already present in systems, avoiding expensive re-indexing or model retraining. The reduced token generation and chunk retrieval also lower API costs for organizations using cloud-based LLMs.
Looking ahead, similar link-aware strategies could extend to other structured formats beyond HTML, including JSON-LD, RDF graphs, and dynamic knowledge bases. The research suggests a broader principle: respecting source document topology during retrieval improves downstream generation quality, potentially influencing how future RAG systems incorporate semantic structure.
- →LARAG leverages existing hyperlink metadata in HTML documentation to improve RAG retrieval without additional graph construction overhead
- →Benchmarking shows LARAG achieves higher answer quality (BERTScore F1) while using fewer retrieved chunks and generated tokens than baseline RAG
- →The approach demonstrates that preserving author-defined link structures significantly improves the semantic grounding of LLM outputs
- →Lightweight link-aware retrieval can be immediately implemented in existing systems without architectural changes or model retraining
- →Results validate that implicit graph-like retrieval through existing metadata outperforms flat-document embedding approaches for technical documentation