ma-finkg 0.1.5

Multi-agent System for Financal Knowledge Graph Construction using LangGraph

MA-FINKG: Multi-Agent Framework for Financial Knowledge Graph Construction from Unstructured Text

A multi-agent system based on LangGraph for constructing Knowledge Graphs from financial text documents. The system uses specialized agents to coordinate domain-specific ontology building, text processing, and knowledge extraction.

Features

• Multi-Agent Architecture: Coordinated agents for different aspects of KG construction
• Benchmarked on varity of datasets: Support for FIRE, NYT11, CoNLLpp, REFinD, and custom ontologies
• Flexible Prompting: Specialized prompt sets optimized for different datasets
• DSPy Integration: Optional optimization for improved extraction performance
• Financial Focus: Specialized for financial text analysis and knowledge extraction

Installation

pip install ma-finkg

Or install from source:

git clone <repository-url>
cd financial_kg_system
pip install -e .

Quick Start

from ma_finkg import MA_FinKG

# Initialize with API key
kg = MA_FinKG(api_key="your-openrouter-key")

# Or set environment variable
import os
os.environ["OPENROUTER_API_KEY"] = "your-key"
kg = MA_FinKG()

# Generate knowledge graph
text = "Apple Inc. reported revenue of $95B in Q4 2023."
result = kg.generate(text)

# Access results
print(f"Found {len(result.entities)} entities and {len(result.triples)} relations")
print("Entities:", result.entities)
print("Relations:", result.triples)

Available Ontologies

FIRE (Default)

Financial dataset ontology with comprehensive financial entities and relations.

Entities: Action, BusinessUnit, Company, Date, Designation, FinancialEntity, GeopoliticalEntity, Location, Money, Person, Product, Quantity, Sector

Relations: ActionBuy, ActionSell, ActionMerge, Employeeof, Subsidiaryof, Value, Locatedin, etc. (18 total)

NYT11

General-purpose ontology based on the NYT dataset.

Entities: PER (Person), ORG (Organization), LOC (Location)

Relations: /people/person/place_of_birth, /business/company/founders, /location/country/capital, etc. (12 total)

Other Ontologies

• CoNLLpp: Named Entity Recognition dataset ontology
• REFinD: ReFiNED dataset ontology
• Default: Generic financial ontology

Configuration Options

kg = MA_FinKG(
    model="openai/gpt-4o-mini",  # or "google/gemini-2.5-flash-lite"
    ontology="fire",             # "fire", "nyt11", "conllpp", "refind", "default"
    prompts="fire",              # matching prompt set
    use_dspy=False,              # True for optimized extraction
    api_key="your-key"           # OpenRouter API key
)

Multi-Agent Architecture

The system employs four specialized agents:

1. KG Expert: Task coordinator that manages the overall KG construction process
2. Domain Specific Expert: Builds domain-specific ontologies from input text
3. Data Processing Expert: Performs lossless text cleaning and preparation
4. Knowledge Extraction Expert: Orchestrates NER and RE sub-agents for entity and relation extraction

Command Line Usage

# Show help
ma-finkg --help
python -m ma_finkg --help

# Show version
ma-finkg --version
python -m ma_finkg --version

Requirements

• Python 3.9+
• langchain-core
• langchain-openai
• langgraph
• dspy-ai
• pydantic
• pyyaml
• json-repair

License

MIT License

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Citation

If you use this software in your research, please cite:

@software{ma-finkg,
  title={MA-FinKG: Multi-Agent System for Financial Knowledge Graph Construction for Text},
  author={Abdel-Rahman Elzohairy},
  version={0.1.5},
  year={2025}
}