KGLite

An embedded Cypher dialect for LLM-agent workloads. A knowledge graph that runs inside your Python process — load data, query with Cypher, hand the graph to an agent via the bundled MCP server. No server to run, no infrastructure to manage, one .kgl file to ship.

The engine is a pure-Rust crate (kglite); the wheel (pip install kglite) is a PyO3 wrapper around it. Bolt and MCP protocol servers are standalone Rust binaries that wrap the same engine. The .kgl file format is portable across all bindings.

Cypher first

Cypher is the primary query surface — agents already know it, and the engine targets an explicitly documented openCypher-compatible subset (including three-valued NULL logic), checked with independently authored local contracts and optional Neo4j differential runs. DataFrame loaders add_nodes() / add_connections() exist to get bulk data in; once it’s in, you query with Cypher.

Embedded, in-process

No server, no network; import and go

LLM-agent surface

Bundled MCP server + describe() schema for system prompts

Cypher subset, honest semantics

Querying + mutations + text_score() for semantic search

In-memory by default

Mapped + disk modes for Wikidata-scale; in-memory is the design centre

Label model

One primary type + optional secondary labels — see multi-label rationale

One-file persistence

.kgl snapshots — copy, share, reload elsewhere

Rust-embeddable

Pure-Rust core; embed without PyO3 — see Rust track

Ecosystem

kglite is the engine. Two companion projects build graphs it serves — each released and versioned on its own cadence:

  • kglite — the embedded Cypher knowledge-graph engine (this project): graph + Cypher + fluent API + bundled MCP server.

  • codingest — parses codebases into code graphs (14 languages, web-framework route detection). Build with it, query the .kgl here. Requires kglite ≥ 0.14.

  • kglite-datasets — fetch-build-cache loaders for public registries (SEC EDGAR, Wikidata, Sodir).

  • sonagram — turns a local music library into a kglite knowledge graph via sonara audio analysis (tempo, energy, mood, key); AI agents curate playlists over it through a simple bundled skill and CLI (pip install sonagram).

Coming from 0.13? The code-graph builder and dataset loaders moved out of the wheel in 0.14 — see the 0.13 → 0.14 migration guide. Pin back anytime with pip install "kglite<0.14".

Pick your track

  • Python guide — pip install kglite, then import kglite. The headline track; covers data loading, Cypher, the MCP server, agents.

  • Rust guide — embed the engine in a Rust binary (cargo add kglite). For graph-as-a-library use cases without the Python wheel.

  • Operators — running the Bolt server (Neo4j wire compat for cluster-aware drivers).

  • Reference — Cypher subset reference + fluent API reference + auto-generated Python API.

  • Concepts — architecture + design decisions + contributor docs.