Build Knowledge Graphs with txt2kg

Basic idea

Transform unstructured text into a structured knowledge graph you can explore and query. This playbook extracts subject–predicate–object triples with a local LLM, stores them in a graph database, and renders the graph in an interactive GPU-accelerated web UI.

The workflow covers:

Knowledge triple extraction — local LLM inference (Ollama or vLLM) to extract relationships from documents
Graph database storage — ArangoDB or Neo4j for storing and traversing triples
GPU-accelerated visualization — Three.js WebGPU for interactive 2D/3D exploration
Web interface — Next.js app for document upload, graph editing, and graph-based queries

What you'll accomplish

A running, containerized system that:

Processes uploaded documents (markdown, text, CSV)
Generates and stores knowledge triples
Lets you visualize and query the graph through a browser

What to know before starting

Required:

Basic Docker container usage
Familiarity with command-line operations

Optional:

Familiarity with knowledge graphs and graph databases

Supported hardware platforms

Use the matrix below to confirm your hardware platform, OS, memory, and whether multi-node applies. The same base workflow applies across supported hardware platforms; ./start.sh always starts the default ArangoDB + Ollama stack.

Hardware platform	OS	Memory	Multi-node capable hardware
DGX Spark	DGX OS (Linux)	128 GB Unified Memory	—
DGX Station	DGX OS (Linux)	Large HBM + Grace DRAM	—

Stack options by hardware platform

Hardware platform	Default start command	Other stack options
DGX Spark	`./start.sh` → ArangoDB + Ollama	`./start.sh --neo4j` → Neo4j + Ollama; `./start.sh --vllm` → Neo4j + vLLM
DGX Station	`./start.sh` → ArangoDB + Ollama	`./start.sh --neo4j` → Neo4j + Ollama; `./start.sh --vllm` → Neo4j + vLLM

IMPORTANT

The 64 KB page-size issue is specific to DGX Station; DGX Spark is not affected. On affected DGX Station systems, prefer ./start.sh --neo4j. Some upstream ArangoDB and Qdrant container images can abort at startup with <jemalloc>: Unsupported system page size; the Neo4j + Ollama stack preserves the fast local Ollama flow while avoiding ArangoDB.

NOTE

Larger models generally produce higher-quality triples. Choose a model that fits the memory available on your hardware platform. See Instructions → Step 3 for defaults and links to explore more models.

Prerequisites

Hardware requirements

Supported hardware platform — see Supported hardware platforms matrix above
Sufficient memory for your chosen LLM

Software requirements

Docker installed and configured with the NVIDIA Container Toolkit
Docker Compose
Network access for container image and model downloads

Ancillary files (in nvidia/playbook-txt2kg/assets after Step 1):

start.sh / stop.sh — launch and shut down services
deploy/compose/ — Docker Compose configurations

Time & risk

Estimated time: 30 MIN (longer on first run while models download; vLLM model load can take 30+ minutes)
Risk level: Low
- GPU memory needs depend on the chosen model
- Document processing time scales with document size and complexity
Rollback: Stop and remove containers; optionally delete downloaded models (see Instructions)
Last Updated: 08/05/2026
- Added explicit Neo4j + Ollama stack option; model defaults and explore links live in Instructions Step 3