[Feature] ArcadeDB backend support

[lvca]

Feature Request: ArcadeDB Backend

Hi Graphiti team,

I'd like to propose adding ArcadeDB as a supported graph database backend.

What is ArcadeDB?

ArcadeDB is an open-source (Apache 2.0), multi-model DBMS that natively supports Graph, Document, Key/Value, Full-Text Search, Time Series, and Vector data — all in a single engine.

Why it's a good fit for Graphiti

Graphiti's hybrid retrieval model requires three things from a backend:

1. Graph traversal — ArcadeDB has a native graph engine with constant-time traversal regardless of database size
2. Vector similarity search — ArcadeDB has native vector indexes (HNSW) for semantic retrieval
3. Full-text search — ArcadeDB has Lucene-based full-text search for BM25 keyword retrieval

ArcadeDB supports all three in a single engine, making it a strong candidate for a self-contained Graphiti deployment.

Implementation note

ArcadeDB 26.2.1 (released February 2026) ships the Neo4j Bolt wire protocol, meaning it accepts connections from standard Neo4j drivers (py2neo, neo4j-python-driver, etc.). The adapter layer from the existing Neo4j backend could serve as a starting point, since the wire protocol is the same.

License: Apache 2.0 (permissive — no BSL or AGPL restrictions)
GitHub: https://github.com/ArcadeData/arcadedb
Docs: https://docs.arcadedb.com
Docker: docker pull arcadedata/arcadedb

I'm happy to help test an integration or provide technical details. Is this something the team would be open to?

[lvca]

Following up on this — I noticed that Kuzu was archived after the Apple acquisition (#1132) and the community is looking for a replacement backend.

I've done some homework on Graphiti's driver architecture (the GraphDriver + 12 operation interfaces pattern under graphiti_core/driver/). A few technical notes on how an ArcadeDB backend would work:

Transport layer: ArcadeDB 26.2.1 ships the Bolt wire protocol, so the neo4j Python async driver connects directly — no custom client needed. The existing Neo4jDriver implementation could serve as the starting template, with the same AsyncGraphDatabase.driver() calls pointed at an ArcadeDB endpoint.

Where it diverges from Neo4j:

• Index creation syntax differs (ArcadeDB uses SQL-based DDL: CREATE INDEX ... ON ... () FULL_TEXT ENGINE LUCENE for full-text, VECTOR engine for HNSW vector indexes)
• Full-text search uses Lucene query syntax via search_indexes() rather than db.index.fulltext.queryNodes()
• Vector search uses vectorNeighbors() function rather than db.index.vector.queryNodes()
• MERGE/MATCH/CREATE Cypher semantics are compatible (OpenCypher TCK-compliant)

What this means practically: the 12 operation implementations would be ~80% shared with Neo4j, with index management and search operations being the main divergence points.

I'd be happy to contribute the driver implementation as a PR. Per the CONTRIBUTING.md, would this need an RFC issue first given it's a new backend (~500+ LOC), or is this feature request sufficient?

[lvca]

Hi maintainers — just following up on my February 25 comment above.

Quick update and a question: I've mapped out the implementation and the Bolt compatibility in ArcadeDB 26.2.1 means the Neo4j async driver connects directly, so the backend adapter would share ~80% of the existing Neo4j backend code. I can have a draft PR ready.

Before I submit, a couple of quick questions:

1. Does the project require an RFC issue before a backend PR of this size (~500 lines)? Happy to open one first if that's the preferred workflow.
2. Is there any concern I should be aware of — licensing, test infrastructure, or API surface — that would affect the contribution?

One more note: I noticed KuzuDB is currently listed as a supported Graphiti backend. Kùzu Inc. was acquired by Apple in October 2025 and the KuzuDB GitHub repo is now archived. For users who chose Kuzu for its embedded, local-first characteristics, ArcadeDB covers the same use case — embedded or server mode, Apache 2.0, with graph + vector + full-text search in a single query.

Happy to coordinate on whatever path works best for the project. Thank you!

[lvca]

Created the driver, it's all in #1310. Please review. Thanks in advance.