ontograph 0.2.0

LLM-powered knowledge graph engine with ontological entity resolution, orbit-based proximity scoring, and hybrid semantic-graph retrieval.

ontograph

LLM-powered knowledge graph engine with ontological entity resolution, orbit-based proximity scoring, and hybrid semantic-graph retrieval.

Install

pip install ontograph
# or from source:
uv add .

Requires OPENAI_API_KEY environment variable.

Quick Start

from ontograph import OntoDB, Schema

# Initialize — creates a local SQLite database
db = OntoDB("my_knowledge.db")

# Define an ontology schema (constrains what the LLM extracts)
db.register_schema(Schema(
    name="workplace",
    entity_types=["person", "project", "team", "meeting", "topic"],
    relationship_types=[
        {"name": "works_on", "directed": True},
        {"name": "manages", "directed": True},
        {"name": "colleague", "directed": False},
        {"name": "discussed", "directed": True},
        {"name": "member_of", "directed": True},
    ],
))

# Ingest unstructured text — entities and relationships extracted automatically
db.ingest(
    "Meeting with Nara and Marco about Project Neptune. "
    "Nara is leading the backend rewrite. Marco raised concerns about the launch deadline.",
    source_type="transcript",
    schema_name="workplace",
)

# Search the graph (semantic + keyword + graph traversal)
results = db.search("What is Nara working on?")

# Ask a question — LLM synthesizes answer from graph context
answer = db.ask("Who is concerned about the launch deadline?")
print(answer)

Core Concepts

Entities, Relationships, Attributes

Everything decomposes into these three primitives:

• Entities: nodes — people, projects, topics, anything nameable
• Relationships: edges — directed (A → B) or bidirectional (A ↔ B)
• Attributes: key-value metadata on entities and relationships

Entity Resolution

When ingesting text, names are fuzzy-matched against existing entities using a composite score:

• Phonetic similarity (Metaphone) — catches pronunciation-similar names
• Spelling similarity (Jaro-Winkler) — catches typos
• Semantic similarity (embedding cosine) — catches conceptual matches
• Orbit proximity — weights by interaction frequency

Orbit

Your "orbit" is a proximity model. Entities you interact with frequently score higher. This powers entity resolution — when "Sal" appears in a transcript, the system weights your manager "Sam" (who you interact with daily) over random "Sal"s elsewhere.

# Add a known transcription error alias
sam = db.get_entity("Sam")
db.add_alias(sam.id, "Sal", alias_type="transcript_error")

Schemas

Ontology schemas define valid entity and relationship types for a domain. They constrain what the LLM extracts.

Self-Improving Feedback Loop

Every entity resolution is logged. Mark resolutions as correct/incorrect to track accuracy over time:

db.mark_resolution("log_id", correct=True)
print(db.stats()["resolution_accuracy"])

API

• OntoDB(db_path, api_key=None, observer_id=None) — Main entry point
• .ingest(text, source_type, schema_name, observer_id, metadata) → dict — Ingest unstructured text
• .search(query, limit, graph_depth) → list[dict] — Hybrid search
• .ask(question) → str — LLM-synthesized answer from graph context
• .add_entity(name, entity_type, attributes, aliases) → Entity — Manual entity creation
• .add_relationship(source, target, type, directed) → Relationship — Manual relationship creation
• .resolve(name, entity_type, observer) → (Entity | None, float) — Entity resolution
• .orbit(observer, limit) → list[dict] — Proximity-ranked entities
• .stats() → dict — Graph statistics and resolution accuracy