Your LLM's new side gig: professional knowledge graph extractor

What it does

kg-gen is a Python library that feeds text to a language model and returns a structured knowledge graph complete with entities, edges, and relations. It handles both short strings and long documents via configurable chunking, plus it can process conversation-style message arrays. The output includes optional clustering maps that group synonymous entities and relations—so “ML” and “machine learning” end up in the same bucket.

The interesting bit

Instead of locking you into a specific model, it routes calls through LiteLLM and uses DSPy for structured outputs, meaning you can point it at OpenAI, local Ollama instances, or Gemini with a single string change. It also treats conversations as first-class citizens, extracting not just factual claims but speaker-to-concept relationships.

Key highlights

• Works with any LiteLLM-supported provider (openai/gpt-4o, ollama_chat/deepseek-r1:14b, etc.)
• Chunks large texts automatically and aggregates multiple graphs into one
• Clusters synonymous entities and relations during or after generation
• Includes an MCP server for AI agent memory integration
• Ships with a built-in visualization tool that exports to HTML

Caveats

• The README shows example outputs but does not include accuracy metrics or benchmark results for its MINE benchmark
• Extraction quality is entirely dependent on the external LLM you configure via LiteLLM

Verdict

Worth exploring if you need knowledge graph extraction for RAG, synthetic data, or conversation analysis and want provider flexibility through LiteLLM. Look elsewhere if you need deterministic parsing independent of LLM inference.