Building the Document Context Layer for AI Agents — Jerry Liu, LlamaIndex

Jerry Liu argues that RAG in 2026 decomposes into an agent harness and a context layer, with retrieval complexity moving into the agent. He introduces a three-layer platform for handling document context: parsing (contrasting heuristic pipelines with one-shot VLM reads and proposing hybrid routing, benchmarked by ParseBench), semantic storage, and repeatable document workflows. He highlights the difficulty of PDF OCR and introduces LightParse, a fast Rust parser for the initial pass before a VLM tool.