growgraph/ontocast: Agentic Ontology Assisted Framework for Semantic Triple Extraction

Agentic Ontology Assisted Framework for Semantic Triple Extraction - growgraph/ontocast

Summary

OntoCast is a Python framework that turns documents into RDF knowledge graphs, building both the ontology and the extracted facts through language models. It runs as an HTTP service, a batch command, or embedded inside LangChain and LangGraph agents, and is licensed under Apache-2.0.

What it is useful for

OntoCast fits work where both the schema and the data need to be constructed from text together. Concrete uses include:

• Pulling RDF triples from PDFs, Office files, HTML, Markdown, or images.
• Maintaining a versioned, evolving ontology alongside the facts it describes, with the model emitting insert/delete patches rather than rewriting the graph.
• Validating output with SHACL shapes and tracing each fact back to its source text via RDF 1.2 provenance.
• Multi-tenant projects using Apache Jena Fuseki for partitioned persistence.
• Augmenting LangChain or LangGraph agents with ontology editing tools so they can edit the schema as the user types.

Entity mentions across a document are merged into a single node, and ontology changes go through a render-and-critique loop before facts are extracted against them.

How engineers can use it

The README documents three modes. Install with the extras you need:

pip install "ontocast[server,openai,doc-processing]"

The openai extra provides a model provider (anthropic, google, and ollama are also listed); doc-processing enables PDF, Office, HTML, Markdown, and image conversion.

Run as an HTTP service:

export LLM_API_KEY=sk-...
ontocast serve
curl -X POST http://127.0.0.1:8999/process -F "[email protected]" -o result.json

The response holds the facts and ontology as Turtle. Configuration is read from environment variables, or from files passed via ontocast --env-file my.env serve (later files override earlier ones).

For batch processing without a server:

ontocast process --input-path ./papers --output-dir ./out

Supply your own ontologies by pointing at a Turtle directory at startup, or uploading to a running server:

ontocast serve --ontology-dir ./my-ontologies
curl -X POST http://127.0.0.1:8999/ontologies -F "[email protected]"

Embed inside a LangChain agent:

from langchain.agents import create_agent
from ontocast import Config, ToolBox, ontocast_tools

tools = await ToolBox.acreate(Config.in_memory())
await tools.initialize()

agent = create_agent(
    model,
    tools=[*ontocast_tools(tools)],
    prompt="Edit the ontology from the user's text.",
)

For LangGraph, make_ontocast_node adds OntoCast as a node, and run_unit_pipeline handles a single passage.

Documented notes and limitations:

• The base install is light and excludes document processing and the ML runtime; both require the relevant extras.
• An LLM API key must be provided at runtime.
• Storage is in-memory by default; persistence requires Apache Jena Fuseki.
• Ontology context can be selected from a catalog, retrieved from a vector store (LanceDB or Qdrant), or fixed.
• Contributors accept a Contributor License Agreement once, on their first pull request.

For configuration reference, the HTTP API, and the Python API, see the documentation at growgraph.github.io/ontocast.