What it does Hyper-Extract is a Python CLI and library that feeds documents to LLMs and returns structured knowledge in eight shapes: simple lists, Pydantic models, graphs, hypergraphs, temporal graphs, spatial graphs, and the full spatio-temporal combo. You pick a YAML template (80+ ship for finance, legal, medical, and general use), point it at a file, and get typed output. It also incrementally evolves knowledge bases as new documents arrive.

The interesting bit The project treats “extraction engine” as a swappable layer. It bundles GraphRAG, LightRAG, Hyper-RAG, KG-Gen, Cog-RAG and others behind a single interface, then maps their outputs onto strongly-typed structures via Pydantic. The hypergraph and spatio-temporal support is genuinely unusual in this space — most tools stop at flat knowledge graphs.

Key highlights

• One CLI command (he parse) from document to graph; he show opens an interactive visualization
• 80+ zero-code YAML templates across six domains, or write your own with a documented schema
• Local deployment via vLLM (tested with Qwen3.5-9B + bge-m3) for air-gapped runs
• Verified against OpenAI, Alibaba Bailian, and local vLLM endpoints; requires json_schema or function-calling support
• Incremental ingestion: feed new docs to expand an existing knowledge base without rebuilding

Caveats

• The comparison table claims features for competitors that are hard to verify (e.g., “Multi-language” checks are based on the author’s own assessment)
• Only three model families are verified; your favorite local model may lack the structured-output capability it needs
• The “10+ extraction engines” include some that are more like configuration variants than distinct algorithms

Verdict Worth a look if you regularly turn messy documents into structured knowledge and want one tool that spans lists to hypergraphs. Skip it if you need battle-tested extraction without LLM hallucination risks, or if your models don’t support strict JSON schema output.