Skip to content

Methods

Guide to extraction algorithms and when to use each.

Overview

Methods are the underlying algorithms that extract knowledge from text. They determine: - How text is processed - How entities and relations are identified - The quality and style of extraction

Method Categories

Typical Methods

Direct extraction without retrieval.

Best for: Smaller documents, direct extraction

graph LR
    A[Document] --> B[LLM Processing]
    B --> C[Extract Structure]

RAG-Based Methods

Retrieval-Augmented Generation combines information retrieval with text generation.

Best for: Large documents, complex queries

graph LR
    A[Document] --> B[Chunk & Index]
    B --> C[Retrieve Relevant]
    C --> D[Generate Answer]
    D --> E[Extract Structure]

Typical Methods

itext2kg

Description: High-quality triple-based extraction

Characteristics:

  • Optimized for triple quality
  • Iterative refinement
  • Good for knowledge abstract construction

Best for:

  • Knowledge graph construction
  • High-quality requirements
  • Triple extraction

Usage:

ka = Template.create("method/itext2kg")

itext2kg_star

Description: Enhanced iText2KG

Characteristics:

  • Improved extraction quality
  • Better handling of complex cases
  • Enhanced entity linking

Best for:

  • When quality is critical
  • Complex extraction scenarios
  • Production systems

Usage:

ka = Template.create("method/itext2kg_star")

kg_gen

Description: Knowledge Graph Generator

Characteristics:

  • Configurable generation
  • Flexible schema
  • Fast processing

Best for:

  • Custom schemas
  • Rapid prototyping
  • Flexible requirements

Usage:

ka = Template.create("method/kg_gen")

atom

Description: Temporal knowledge graph with evidence

Characteristics:

  • Temporal fact extraction
  • Evidence attribution
  • Confidence scoring

Best for:

  • Temporal analysis
  • Fact verification
  • Timeline extraction

Usage:

ka = Template.create("method/atom")

RAG-Based Methods

light_rag

Description: Lightweight Graph-based RAG

Characteristics:

  • Fastest RAG method
  • Binary edges (source → target)
  • Good balance of speed/quality
  • Suitable for most use cases

Best for:

  • General-purpose extraction
  • Medium to large documents
  • Quick results

Usage: 

ka = Template.create("method/light_rag")

graph_rag

Description: Graph-RAG with Community Detection

Characteristics:

  • Community detection for organization
  • Hierarchical summaries
  • Best for very large documents
  • Slower but thorough

Best for:

  • Very large documents (books, long papers)
  • Complex topic structures
  • Research papers

Usage: 

ka = Template.create("method/graph_rag")

hyper_rag

Description: Hypergraph-based RAG

Characteristics:

  • N-ary hyperedges (2+ entities)
  • Captures complex relationships
  • Richer graph structure

Best for:

  • Multi-party relationships
  • Project collaborations
  • Complex organizational structures

Usage: 

ka = Template.create("method/hyper_rag")

hypergraph_rag

Description: Advanced Hypergraph RAG

Characteristics:

  • Enhanced hypergraph capabilities
  • Advanced relationship modeling

Best for:

  • Complex hypergraph scenarios
  • Advanced relationship analysis

Usage:

ka = Template.create("method/hypergraph_rag")

cog_rag

Description: Cognitive RAG

Characteristics:

  • Cognitive retrieval mechanisms
  • Reasoning-focused

Best for:

  • Reasoning tasks
  • Question-answering systems

Usage:

ka = Template.create("method/cog_rag")

Selection Guide

By Document Size

Size Recommended
Small (< 1K words) itext2kg, kg_gen
Medium (1-10K) light_rag, itext2kg_star
Large (10-50K) light_rag, graph_rag
Very Large (> 50K) graph_rag

By Use Case

Use Case Recommended
Quick extraction light_rag
Best quality itext2kg_star
Large documents graph_rag
Complex relationships hyper_rag
Temporal facts atom
Knowledge bases itext2kg

By Priority

Priority Recommended
Speed light_rag
Quality itext2kg_star
Cost light_rag
Completeness graph_rag

Comparison Table

Method Speed Quality Memory Best For
itext2kg ⭐⭐⭐ ⭐⭐⭐ ⭐ Quality focused
itext2kg_star ⭐⭐⭐ ⭐⭐⭐ ⭐ Production quality
kg_gen ⭐⭐⭐ ⭐⭐ ⭐ Flexibility
atom ⭐⭐ ⭐⭐⭐ ⭐⭐ Temporal data
light_rag ⭐⭐⭐ ⭐⭐ ⭐⭐ General use
graph_rag ⭐ ⭐⭐⭐ ⭐⭐⭐ Large docs
hyper_rag ⭐⭐ ⭐⭐⭐ ⭐⭐⭐ Complex relations

Listing Available Methods

from hyperextract.methods import list_methods

methods = list_methods()
for name, info in methods.items():
    print(f"{name}: {info['description']}")
    print(f"  Type: {info['type']}")

See Also