An academic extraction toolkit that grew into an LLM swiss army knife

What it does DeepKE is an open-source toolkit for constructing knowledge graphs through information extraction. It handles named entity recognition, relation extraction, attribute extraction, and event extraction across both standard supervised and low-resource settings. Originally built around fine-tuned deep learning models, the project now also hosts LLM-based extraction via DeepKE-LLM and OneKE, plus MCP service tools that let large models call lightweight extraction utilities.

The interesting bit Rather than abandoning its pre-LLM architecture, DeepKE layers new capabilities on top: you can still run LightNER or KnowPrompt for efficient few-shot extraction, or swap in a 13B instruction-tuned model for zero-shot work. The recent MCP integration effectively turns the toolkit into a set of callable tools for LLM agents, which is a pragmatic admission that small specialized models and large general ones each have their place.

Key highlights

• Supports Chinese and English extraction, including a schema-based bilingual model (OneKE) and instruction datasets (IEPile, InstructIE).
• Covers the full pipeline: NER, relation extraction, attribute extraction, and event extraction in one framework.
• Ships with off-the-shelf models under DeepKE-cnSchema for users who want pretrained weights rather than training from scratch.
• Implements both document-level and multimodal extraction scenarios, not just sentence-level text.
• Publishes its own EMNLP 2022 paper and maintains active documentation, demo sites, and Colab notebooks.

Caveats

• The README warns that Linux is strongly recommended and Windows path handling can be brittle.
• Dependency constraints are strict: for example, the traditional pip install deepke path requires pip version ≤ 24.0, while the LLM branch needs a completely separate environment.
• The codebase spans multiple eras—classical DeepKE, DeepKE-LLM, and MCP tools—so navigating the repository can feel like reading a palimpsest.

Verdict Researchers and engineers working on Chinese-English knowledge graph construction should look here, especially if they want to compare classical and LLM-based extraction without switching frameworks. If you need a single, opinionated pipeline with clear versioning, the sprawling scope may frustrate you.