unstructured2graph

Convert unstructured documents into knowledge graphs within Memgraph.

Overview

unstructured2graph enables you to transform any unstructured data (PDFs, URLs, documents) into a graph database, powering Graph Retrieval-Augmented Generation (GraphRAG) applications. It combines:

• Unstructured - Parse and chunk diverse document formats
• LightRAG - Extract entities and relationships using LLMs
• Memgraph - Store and query your knowledge graph

Installation

Install from source:

git clone https://github.com/memgraph/ai-toolkit.git
cd ai-toolkit/unstructured2graph
pip install -e .

For full document support (PDF, DOCX, etc.):

pip install -e ".[all-docs]"

Quick Start

import asyncio
from memgraph_toolbox.api.memgraph import Memgraph
from lightrag_memgraph import MemgraphLightRAGWrapper
from unstructured2graph import from_unstructured, create_property_index


async def main():
    memgraph = Memgraph(user_agent="unstructured2graph")
    create_property_index(memgraph, "Chunk", "hash")

    lightrag = MemgraphLightRAGWrapper()
    await lightrag.initialize(working_dir="./lightrag_storage")

    # Ingest documents from URLs or local files
    await from_unstructured(
        sources=["https://example.com/doc.pdf", "./local_file.md"],
        memgraph=memgraph,
        lightrag_wrapper=lightrag,
        link_chunks=True,  # Create NEXT relationships between chunks
    )
    await lightrag.afinalize()


asyncio.run(main())

Persistence: MemgraphLightRAGWrapper now persists LightRAG's full working state into Memgraph by default — the entity/relationship graph plus the key/value store, vector store, and document-status store. The working_dir argument is still accepted (and used as a fallback location for any store not backed by Memgraph), but with the default settings the JSON stores are no longer written there. See the lightrag-memgraph README for the label/index schema and opt-out flags.

Key Features

API Reference

Document Processing

• parse_source(source, partition_kwargs) - Parse a single file or URL into chunks
• parse_text(text, partition_kwargs) - Chunk a raw in-memory string (no file/URL involved)
• make_chunks(sources, partition_kwargs) - Process multiple sources into ChunkedDocument objects
• from_unstructured(sources, memgraph, lightrag_wrapper, ...) - Full ingestion pipeline for files/URLs; returns one Chunk group per source
• from_texts(texts, memgraph, lightrag_wrapper, ...) - Full ingestion pipeline for raw strings; returns one Chunk group per input text

Graph Operations

• create_nodes_from_list(memgraph, nodes, label, batch_size) - Batch insert nodes
• connect_chunks_to_entities(memgraph, chunk_label, entity_label) - Link entities to source chunks
• link_nodes_in_order(memgraph, ...) - Create sequential relationships between chunks
• create_vector_search_index(memgraph, label, property) - Create vector index for similarity search
• compute_embeddings(memgraph, label) - Generate embeddings for nodes

Documentation

For detailed usage examples and getting started guides, check out the official documentation:

👉 unstructured2graph Documentation

Requirements

• Python 3.10+
• Memgraph database instance

LLM API Key

This library uses LightRAG for entity and relationship extraction, which requires an LLM API key. Set your OpenAI API key as an environment variable:

export OPENAI_API_KEY="your-api-key"