Runic

A type-safe OGM for Cypher graph databases.
Define your models once, run them on any backend.

Why Runic • The OGM • Migrations • Graph-RAG • Installation • Docs

Runic maps Python classes to graph nodes and edges. You declare typed Node and Edge models and get change tracking, lazy and eager relationships, a composable query API, and schema migrations — on top of a pluggable driver layer that runs the same model code on FalkorDB, Neo4j, Memgraph, ArcadeDB, Apache AGE, and Amazon Neptune (Database and Analytics). A built-in Graph-RAG SDK layers document ingestion and cited, knowledge-graph-grounded question answering on top of the OGM.

Why Runic

• Backend-agnostic. One model definition runs on seven backends. Switching from FalkorDB to Neo4j means changing the arguments to create_driver(); your models, queries, and application code don't change.
• Typed models, no metaclass magic. Node and Edge are plain classes with typed Field descriptors — IDE autocomplete works, and Author.name == "Alice" builds a query predicate.
• Change tracking. Mutate an object and call commit(); the unit-of-work session computes the diff and writes only what changed. No manual dirty flags, no hand-written SET clauses.
• First-class relationships. Declare Relation fields for INCOMING/OUTGOING edges, with edge-property models, and choose lazy loading or single-round-trip eager fetch.
• Native graph types. Vector (vecf32), GeoLocation (point), interned strings, and automatic converters for datetime and Enum — stored without writing serialization code.
• Migrations included. A migration tool with versioned revisions, a CLI, and rollback snapshots. Revision state lives inside the graph, so there's no external state table.
• Sync and async. Session/Repository and AsyncSession/AsyncRepository share one API.

The OGM

The two examples below build on one domain — Author, Article, and an AUTHORED edge — and cover the features you'll reach for most.

1. Model a domain, persist it, and wire up relationships

Declare typed Node and Edge classes. Constraints and indexes go inline on Field; native graph types (Vector, GeoLocation, datetime, Enum) get their converters assigned automatically; and Relation declares a traversal with an optional edge-property model. SchemaManager reconciles those declarations with the live graph, a Session handles writes with automatic change tracking, and session.relate() creates the edges between nodes.

from datetime import UTC, datetime
from enum import StrEnum

from runic.ogm import (
    Edge,
    Field,
    GeoLocation,
    Node,
    Relation,
    SchemaManager,
    Session,
    Vector,
    create_driver,
)


class Status(StrEnum):
    DRAFT = "draft"
    PUBLISHED = "published"


class Article(Node, labels=["Article"]):
    id: str = Field(primary_key=True)
    title: str = Field(
        index_type="FULLTEXT"
    )  # fulltext search: FalkorDB/Neo4j/Memgraph
    category: str = Field(
        interned=True
    )  # intern() dedup — FalkorDB only, no-op elsewhere
    status: Status = Status.DRAFT  # EnumConverter auto-assigned
    published_at: datetime | None = None  # DatetimeConverter auto-assigned
    embedding: Vector | None = None  # KNN via vecf32() on FalkorDB; Neo4j/Memgraph
    #   need a pre-created VECTOR INDEX
    origin: GeoLocation | None = None  # point(); updates unsupported on ArcadeDB


class AuthoredEdge(Edge, type="AUTHORED"):
    created_at: datetime


class Author(Node, labels=["Author"]):
    id: str = Field(primary_key=True)
    email: str = Field(unique=True)
    name: str
    articles: list[Article] = Relation(
        relationship="AUTHORED",
        direction="OUTGOING",
        target="Article",
        edge_model=AuthoredEdge,
    )


# Pick a backend here — nothing else in this file changes.
driver = create_driver("falkordb", host="localhost", port=6379, graph="blog")

# Reconcile declared indexes/constraints with the live graph.
schema = SchemaManager(driver)
schema.sync_schema([Author, Article], drop_extra=False)  # create missing; keep extras

with Session(driver) as session:
    alice = Author(id="alice", email="alice@example.com", name="Alice")
    intro = Article(
        id="a1", title="Graphs 101", category="intro", status=Status.PUBLISHED
    )
    session.add_all([alice, intro])
    session.commit()

    # Create the AUTHORED edge, writing properties onto the relationship itself.
    # relate() is MERGE-based: idempotent, and re-calling updates the edge props.
    session.relate(
        alice, Author.articles, intro, edge=AuthoredEdge(created_at=datetime.now(UTC))
    )
    session.commit()

with Session(driver) as session:
    alice = session.get(Author, "alice")
    alice.name = "Alice Smith"  # tracked automatically — no explicit dirty flag
    session.commit()  # only the diff is written

2. Query and traverse the graph

Read data back with composable, type-safe statements, multi-hop traversals, paginated repositories, or your own Cypher. select() builds a statement independently of any session, so you can assemble it from conditional filters and reuse it across sessions. .traverse() walks a relationship — one hop, or to any depth with hops= — for real-world graph queries.

from runic.ogm import Repository, Session, alias, select


# Compose a query dynamically, then run it three ways.
stmt = select(Article).where(Article.status == Status.PUBLISHED)
if category:
    stmt = stmt.where(Article.category == category)

with Session(driver) as session:
    articles: list[Article] = session.scalars(stmt)  # list[Article]
    latest: Article | None = session.scalar(stmt)  # Article | None
    n: int = session.count(stmt)  # int

    # Single-hop traversal with an edge-property filter — published articles
    # Alice authored after a cutoff date.
    e, art = alias(AuthoredEdge, "e"), alias(Article, "art")
    recent = (
        session.query(Author, "a")
        .where(Author.id == "alice")
        .traverse(Author.articles, to=art, edge=e)
        .where(e.created_at >= cutoff)
        .where(art.status == Status.PUBLISHED)
        .return_target(art)
        .all()
    )

    # Variable-length traversal — every article reachable within 3 AUTHORED hops
    # (e.g. co-authorship chains). hops=(1, None) means unbounded.
    network = (
        session.query(Author, "a")
        .where(Author.id == "alice")
        .traverse(Author.articles, to="reached", hops=(1, 3))
        .all()
    )

    # Paginate through a repository.
    page = Repository(session, Article).find_all(skip=0, limit=20)

    # Or load relationships off an entity: lazy by default, eager on request.
    author = session.get(Author, "alice")
    author.articles  # lazy — queried on first access
    eager = session.get(Author, "alice", fetch=["articles"])
    eager.articles  # already loaded, no extra query


# Subclass Repository to drop down to typed Cypher when you need it.
class ArticleRepository(Repository[Article]):
    def by_author_email(self, email: str) -> list[Article]:
        return self.cypher(
            "MATCH (:Author {email: $email})-[:AUTHORED]->(a:Article) RETURN a",
            {"email": email},
            returns=Article,
        )

Migrations

runic.migrate is a migration tool with a CLI for versioned schema evolution. It stores revision state inside the graph, so there's no external state table to manage.

runic init
runic revision -m "create user index"

Edit the generated file in runic/versions/:

revision = "1975ea83b712"
down_revision = None


def upgrade(op) -> None:
    op.create_range_index("User", "email")


def downgrade(op) -> None:
    op.drop_range_index("User", "email")

Apply or roll back:

runic upgrade            # apply all pending revisions
runic downgrade          # roll back one step
runic downgrade 1975e    # roll back to a specific revision (prefix is enough)

Baseline an existing graph without re-running anything — introspect, generate a root revision, and stamp it. The generated revision rebuilds the full schema on an empty graph, so it's safe to replay for CI, cloning, or new tenants:

runic baseline -m "baseline"   # introspect, generate root revision, stamp it
runic current                  # verify it is now tracked
runic upgrade head             # rebuild full schema on a fresh graph

Programmatic SDK — drive migrations from code, against any backend:

from pathlib import Path
from runic import Runic, init
from runic.migrate.adapters import create_adapter

init(Path("runic/"))

adapter = create_adapter(
    "falkordb", url="falkor://localhost:6379", graph_name="my_graph"
)
# adapter = create_adapter("neo4j", host="localhost", port=7687,
#                          database="neo4j", username="neo4j", password="secret")

runic = Runic(adapter, script_location=Path("runic/"))
runic.migrate.upgrade("head")
print("current:", runic.migrate.current())

Graph-RAG

runic.rag turns unstructured text into a knowledge graph and answers questions over it with citations — chunking, entity/relation extraction, embedding, storage, and hybrid (vector + fulltext + graph-expansion) retrieval are handled for you. It builds on the same OGM and driver layer, so it runs on any supported backend. OpenAI is the default for the LLM and embeddings; point it at Ollama for a fully local run.

from runic.ogm import create_driver
from runic.rag import GraphRAG, Ontology, RagSettings

settings = RagSettings()
driver = create_driver(
    "falkordb", host="localhost", port=6379, graph=settings.falkordb_graph
)

# with_defaults() wires the full adapter stack from your environment:
# paragraph chunker, extractor, embedder, resolver, retrievers, reranker, synthesizer.
rag = GraphRAG.with_defaults(driver, settings=settings, ontology=Ontology.default())
rag.bootstrap_schema()  # create entity types + indexes (idempotent)

# Ingest: chunk -> extract entities/relations -> embed -> resolve -> write the graph.
rag.ingest_text(
    "Ada Lovelace worked with Charles Babbage on the Analytical Engine in London.",
    source="inline-demo",
)

# mode="auto" classifies the question and picks the retrieval strategy automatically.
answer = rag.query("Who worked on the Analytical Engine?")
print(answer.text)
for citation in answer.citations:
    print(f"  - [{citation.source}] {citation.text[:80]}...")

Every stage is a port with a default adapter you can swap — custom ontologies, chunkers, extractors, or retrievers — and the optional runic-rag-docling add-on plugs Docling in for layout-aware parsing and chunking of PDFs and office documents.

uv add "runic-py[graphrag,falkordb]"   # Graph-RAG extras + a backend driver

Installation

Install the core package plus the extra for your backend. The core has no graph-driver dependency — you only pull in what you use.

uv add "runic-py[falkordb]"   # FalkorDB
uv add "runic-py[neo4j]"      # Neo4j
uv add "runic-py[memgraph]"   # Memgraph (Bolt)
uv add "runic-py[arcadedb]"   # ArcadeDB (Bolt)
uv add "runic-py[age]"        # Apache AGE (PostgreSQL extension)
uv add "runic-py[neptune]"    # Amazon Neptune Database (Bolt + IAM auth)
uv add "runic-py[neptune-analytics]"   # Amazon Neptune Analytics (HTTPS)
uv add "runic-py[all]"        # everything
uv add "runic-py[graphrag,falkordb]"   # Graph-RAG SDK + a backend driver

Documentation

Full conceptual overview, async usage, advanced CLI flags, the Graph-RAG guides, and the complete API reference live at the Runic Documentation.

License

Released under the MIT License.

Disclaimer

The code and the documentation are generated with AI.