OntoCast

Agentic ontology assisted framework for semantic triple extraction from documents

Overview

OntoCast is a powerful framework that automatically extracts semantic triples from documents using an agentic approach. It combines ontology management with natural language processing to create structured knowledge from unstructured text.

Key Features

• Ontology-Guided Extraction: Uses ontologies to guide the extraction process and ensure semantic consistency
• Entity Disambiguation: Resolves entity and property references across chunks
• Multi-Format Support: Handles various input formats including text, JSON, PDF, and Markdown
• Semantic Chunking: Intelligent text chunking based on semantic similarity
• MCP Compatibility: Fully compatible with the Model Control Protocol (MCP) specification, providing standardized endpoints for health checks, info, and document processing
• RDF Output: Generates standardized RDF/Turtle output

Extraction Steps

• Document Processing

  • Supports PDF, markdown, and text documents
  • Automated text chunking and processing

• Automated Ontology Management

  • Intelligent ontology selection and construction
  • Multi-stage validation and critique system
  • Ontology sublimation and refinement

• Knowledge Graph Integration

  • RDF-based knowledge graph storage
  • Triple extraction for both ontologies and facts
  • Configurable workflow with visit limits
  • Chunk aggregation preserving fact lineage

Installation

uv add ontocast 
# or
pip install ontocast

Configuration

Create a .env file with your OpenAI API key:

cp .env.example .env

Running the Server

uv run serve \
    --ontology-directory ONTOLOGY_DIR \
    --working-directory WORKING_DIR

Process Endpoint

The /process endpoint accepts: - application/json: JSON data - multipart/form-data: File uploads

And returns: - application/json: Processing results including: - Extracted facts in Turtle format - Generated ontology in Turtle format - Processing metadata

# Process a PDF file
curl -X POST http://url:port/process -F "file=@data/pdf/sample.pdf"

curl -X POST http://url:port/process -F "file=@test2/sample.json"

# Process text content
curl -X POST http://localhost:8999/process \
    -H "Content-Type: application/json" \
    -d '{"text": "Your document text here"}'

MCP Endpoints

OntoCast implements the following MCP-compatible endpoints:

• GET /health: Health check endpoint
• GET /info: Service information endpoint
• POST /process: Document processing endpoint

Processing Filesystem Documents

uv run serve \
    --ontology-directory ONTOLOGY_DIR \
    --working-directory WORKING_DIR \
    --input-path DOCUMENT_DIR

NB

• json documents are expected to contain text in text field
• recursion_limit is calculated based on max_visits * estimated_chunks, the estimated number of chunks is taken to be 30 or otherwise fetched from .env (vie ESTIMATED_CHUNKS)
• default 8999 is used default port

Docker

To build docker

docker buildx build -t growgraph/ontocast:0.1.1 . 2>&1 | tee build.log

Project Structure

src/
├── agent.py          # Main agent workflow implementation
├── onto.py           # Ontology and RDF graph handling
├── nodes/            # Individual workflow nodes
├── tools/            # Tool implementations
└── prompts/          # LLM prompts

Workflow

The extraction follows a multi-stage workflow:

1. Document Preparation

   • [Optional] Convert to Markdown
   • Text chunking

2. Ontology Processing

   • Ontology selection
   • Text to ontology triples
   • Ontology critique

3. Fact Extraction

   • Text to facts
   • Facts critique
   • Ontology sublimation

4. Chunk Normalization

   • Chunk KG aggregation
   • Entity/Property Disambiguation

5. Storage

   • Knowledge graph storage

Documentation

Full documentation is available at: growgraph.github.io/ontocast

Roadmap

1. Add a triple store for serialization/ontology management
2. Replace graph to text by a symbolic graph interface (agent tools for working with triples)

Contributing

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

Acknowledgments

• Uses RDFlib for semantic triple management
• Uses docling for pdf/pptx conversion
• Uses OpenAI language models / open models served via Ollama for fact extraction
• Uses langchain/langgraph