Retrieval-Augmented Generation (RAG) systems face challenges with temporal queries that span multiple documents, such as tracking entity history or the evolution of company strategies over years. The LLM-Wiki architecture addresses this by compiling extensive semantic knowledge during an ingestion phase, creating canonical pages linked by an index. This approach moves reasoning to ingestion time, extracting concepts, entities, and relationships from incoming documents to update persistent knowledge bases on topics like acquisitions or AI strategy. For instance, multiple annual reports would contribute to a consolidated "Acquisitions" page.

However, LLM-Wiki's preemptive semantic compilation incurs significant ingestion costs and faces limitations. As the number of extraction objectives grows, LLMs can suffer from "high-dimensional information extraction degradation," reducing recall and missing salient details, especially in long documents. A fundamental problem is anticipating all future queries; materializing every potentially relevant detail risks creating an unwieldy, costly knowledge base, while omitting details necessitates revisiting original documents. This creates a trade-off between completeness, maintainability, and ingestion cost.

In contrast, the Proxy-Pointer (PP) architecture offers a "lazy" semantic analysis approach. At ingestion, PP builds a purely regex-based structural representation (a skeletal tree) of each document with zero LLM cost. Semantic compilation is deferred to retrieval time. PP modifies standard vector indexes by chunking data within section boundaries, tagging chunks with metadata pointing to their section and line numbers. When a query arrives, a Vector search plus LLM re-ranker pipeline identifies relevant top-k sections, using the document's structure as a semantic map for precise retrieval. This method aims for precision without the "ingestion tax" associated with full semantic precompilation.

Proxy-Pointer's structure-aware retrieval is particularly effective for cross-document queries that require correlating information across multiple sources, such as tracking a company acquired years ago through various annual reports. By preserving structural organization and performing semantic synthesis only when needed, PP avoids the upfront computational burden and predictive challenges of LLM-Wiki's knowledge compilation, making it a more cost-efficient alternative for complex, temporal information retrieval.