Hyper-Extract turns any document into a knowledge graph with one CLI command

Yifan Feng released Hyper-Extract, an open-source command-line tool that turns unstructured text into structured knowledge using LLMs. The framework supports eight knowledge structures — from simple lists to hypergraphs and spatio-temporal graphs — and ships with more than 80 pre-built YAML templates across finance, legal, medical, and general domains.

The project is notable not for a single breakthrough but for how it packages capabilities that have been scattered across separate tools. GraphRAG handles knowledge graphs. LightRAG does lightweight graph extraction. KG-Gen generates graphs from text. Hyper-Extract wraps all three — plus Hyper-RAG, Cog-RAG, and several others — behind a single CLI command: he parse. The user picks a template, points it at a document, and gets back a structured knowledge base that can be queried with natural language or visualized as an interactive graph.

What makes Hyper-Extract different

The comparison table on the project page is the clearest signal of intent. It lists five tools — GraphRAG, LightRAG, KG-Gen, ATOM, and Hyper-Extract — across seven features. Hyper-Extract is the only one that supports spatial graphs, hypergraphs, domain templates, an interactive CLI, and multi-language extraction. ATOM comes closest but lacks spatial graphs and hypergraphs. GraphRAG and LightRAG, the most widely used open-source graph RAG tools, each miss at least three features.

This is not a claim that Hyper-Extract outperforms those tools on quality. The project does not publish benchmark numbers. The claim is combinatorial: Hyper-Extract offers more formats and more templates in one install than any alternative. For a researcher or analyst who needs to extract entities from a 20-page paper, then query the results, then visualize them, the alternative has been to stitch together GraphRAG, NetworkX, a vector database, and a lot of glue code. Hyper-Extract collapses that into he parse paper.pdf -t general/academic_graph -o ./paper_kb/ and he show ./paper_kb/.

The template system is the moat

The 80-plus YAML templates are the project’s strongest feature. Each template defines a schema — entity types, relation types, field names, and identifiers — that the LLM fills during extraction. The academic graph template, for example, defines entities with name, type, and description fields, and relations with source, target, and type. The LLM receives the template as a structured output specification and returns JSON that matches it exactly.

This is not novel in isolation. Pydantic models and JSON schema have been used for structured extraction since GPT-4’s function calling launched. What Hyper-Extract does is pre-package the schemas into domain-specific templates and make them discoverable. A financial analyst extracting earnings reports does not need to design a schema for companies, executives, and financial metrics. They run he parse earnings.md -t finance/earnings_graph. The template handles the schema design. The user handles the document.

The templates are also extensible. Users can write custom YAML templates or modify existing ones. The project’s Apache-2.0 license makes commercial reuse straightforward.

Local deployment and the vLLM integration

Hyper-Extract supports local deployment through vLLM, which means data never leaves the machine. The project provides a code example using Qwen3.5-9B with GPTQ-Marlin quantization and bge-m3 for embeddings. This matters for regulated industries — healthcare, legal, finance — where sending documents to OpenAI’s API is not an option.

The local setup is not trivial. It requires running two vLLM instances: one for the LLM and one for the embedder. The documentation covers the configuration, but the user needs to manage GPU memory, model downloads, and endpoint health. Hyper-Extract does not abstract away the infrastructure; it provides the client configuration once the infrastructure is running.

Still, the option exists. Most knowledge extraction tools that offer this range of output formats are cloud-only. Hyper-Extract gives the user a choice between OpenAI, Alibaba Cloud’s Bailian, and local vLLM. The embedder support mirrors the LLM support: text-embedding-3-small, text-embedding-v4, or bge-m3.

What this means for AI builders

The knowledge extraction space has been fragmented. Researchers who want to turn papers into graphs use GraphRAG. Analysts who want temporal extraction use ATOM. Anyone who wants hypergraphs writes custom code. Hyper-Extract does not solve the hard problems — it does not improve extraction accuracy, it does not handle ambiguous entity resolution, it does not guarantee consistency across documents. What it does is lower the integration cost.

For an AI builder evaluating whether to use knowledge graphs in a product, Hyper-Extract reduces the evaluation cycle from weeks to hours. Install the tool, point it at a sample document, inspect the output, query it, visualize it. If the quality is sufficient, the builder can then invest in customizing templates, tuning the LLM, and building the production pipeline. If not, they move on without having built a custom extraction pipeline.

The project is also a signal of where the open-source LLM tooling ecosystem is heading. The barrier to building a useful CLI tool that wraps multiple LLM backends and multiple extraction algorithms is dropping. Hyper-Extract is one developer’s project, built on top of OpenAI-compatible APIs and open-source models. It does not require a team or a company. It requires a clear abstraction — the template system — and the discipline to ship 80 presets.

The open question

Hyper-Extract does not publish accuracy numbers. The README shows example outputs and comparison tables, but no precision, recall, or F1 scores against standard benchmarks. For a tool that claims to turn “unstructured text into structured knowledge,” the absence of evaluation is a gap. Users must evaluate quality themselves, document by document, template by template.

The project also depends entirely on the LLM’s structured output capability. If the underlying model produces malformed JSON or hallucinates entities, the graph will be wrong. Hyper-Extract does not validate or correct the LLM’s output beyond schema conformance. The quality ceiling is the quality of the model, not the tool.

Those caveats aside, Hyper-Extract is the most complete open-source packaging of LLM-based knowledge extraction available today. It does not invent a new extraction algorithm. It invents a new interface to existing algorithms, and that interface is a single command.