Chunk the document (500–1,000 token chunks with 100–200 token overlap) and store embeddings + metadata (doc ID, section, offsets). Create section-level summaries and an index map; run hierarchical summarization (section → chapter → full) instead of loading the whole doc at once. For quotes, always retrieve the original chunk and paste the exact excerpt with its source ID—don’t rely on the model’s memory. Use a long-context model like Claude or a RAG pipeline for best attribution.

Compare Claude and ChatGPT