knowledge-graph-extractlisted

# Knowledge Graph Extraction Extract subject–relation–object triplets from documents into an append-only JSONL dataset, with deterministic chunking and a progress manifest so an interrupted run resumes exactly where it stopped. Optional schema enforcement and Cypher generation via bundled stdlib-only Python scripts (no pip dependencies, no network, no database connections). **Core principle: the output directory is the state.** Every step reads its inputs from and records its result in `<output_dir>` before moving on. If the session dies at any point, re-invoking this skill on the same directory continues the work. File formats: see [references/output-format.md](references/output-format.md). Schema format: see [references/schema-format.md](references/schema-format.md). ## When not to use - **Very large corpora where token cost dominates.** Every chunk re-reads the entity registry; across thousands of documents that adds up. A dedicated pipeline (LangChain, LlamaIndex, GraphRAG) may be cheaper when you don't need agent-quality reading of mixed formats. - **When you need measured extraction quality.** There is no evaluation harness here; precision/recall rest on the agent's judgment. For benchmarked extraction, use a system with a labelled eval set. ## Phase 0 — Inputs Collect from the user (ask only for what's missing): 1. **Documents**: a folder (recurse; skip hidden files and files you cannot read) or an explicit file list. Any format you can Read works — n