Compare long-context LLMs with the dual-layer GraphRAG pipeline

Determine whether a modern long-context large language model given the full corpus achieves performance comparable to the vector-RAG and Lexical_ReAct systems on small- to moderate-sized CMC document collections.

Background

The paper proposes a dual-layer GraphRAG architecture for CMC process-development documents and argues that its principal advantages emerge for multi-hundred-document programs, where flat retrieval faces cost, scale, structural-reasoning, and update challenges.

For the 52,000-word testbed of 38 documents, the authors suggest that providing the entire corpus to a modern long-context LLM could produce results comparable to those of the vector-RAG baseline or Lexical_ReAct. However, no such comparison is conducted, so the relative value of GraphRAG in small- to moderate-sized corpora remains unresolved.

References

Indeed, for a testbed of 52{,}000 (38 documents), a long-context LLM given the full corpus might score comparably to either the vector-RAG or Lexical_ReAct. While the comparison is not tested here, we argue that the graph's motivating regime is found in multi-hundred-document programs mentioned in \autoref{sec:corpus}.

— From Document Silos to Process Intelligence: A Multi-Layer Knowledge Graph for CMC Process Development  (2609.11493 - Amirmoshiri et al., 10 Sep 2026) in Section 4, Summary, Limitations and Outlook, subsection “Large Context LLM vs GraphRAG”

Retrieval methods can reduce context length and inference cost, but their cost–performance tradeoffs in realistic clinical settings remain unclear [5, 30, 43].

— A Living Benchmark for Information Retrieval from Electronic Health Records  (2609.30205 - Cahoon et al., 24 Sep 2026) in Discussion