Looking for the Python package? This page documents the Java port on Maven Central, which is early and partial (see "Not ported yet"). The Python library has its own reference page.

47 tests in the 0.1.0 release build (0 failing, 13 skipped without their environment) · live-verified on Neo4j 5.26.29 Community only

ragleap-graph (Java)

An early, partial Java port of the ragleap-graph Neo4j graph index, published to Maven Central. Regex entity extraction, document upsert, and relation, lineage and entity queries over the Neo4j HTTP Query API.

Install

<dependency>
  <groupId>io.github.antonyrag</groupId>
  <artifactId>ragleap-graph</artifactId>
  <version>0.1.0</version>
</dependency>

Requires Java 21. The only runtime dependency is jackson-databind. You need a Neo4j server with the HTTP Query API; only Neo4j 5.26 Community has been tested.

What is in 0.1.0

Entity-name normalization, regex entity extraction and composite keys, with Python-compatible string handling
GraphIndex.upsertDocument: the full write path, with retries on transient errors, old-relation cleanup and an optional audit hook
Queries: findDocumentsByEntities, documentEntities, searchRelatedEntities, findRelations, findLineage and findEntitiesByType
Talks to Neo4j over the HTTP Query API v2 through a small CypherRunner interface, with no Bolt driver. GraphConfig takes an HTTP URL, not a Bolt URI. If the server cannot be reached, methods return empty results, like the Python package

Types in the package

The package com.ragleap.graph contains GraphIndex, GraphConfig, CypherRunner, QueryApiRunner, AuditSink, UpsertPlanner, UpsertSummary, EntityExtraction, EntityRef, TypedEntity, CompositeKey, PyText, DocumentHit, RelatedEntity, RelationHit, LineageEntry, GraphException and GraphTransientException. Signatures are in the source.

How it was verified

Compared with the real Python ragleap-graph 0.9.0, pinned at commit 226fa4d: 252 pure-logic cases, 14 upsert scenarios and 47 query scenarios. For the scenarios, the Cypher statements, parameters (int versus float), return values and audit events had to match. A mutation check confirmed the parity tests fail when behaviour differs
47 tests in the release build: 0 failures, 13 skipped (the parity and live tests skip when their environment variables are unset, as in CI)
Live tests passed against Neo4j 5.26.29 Community
After release, the published jar was resolved from Maven Central in a clean project and one document was upserted into a live Neo4j

Limits

Verified only on Neo4j 5.26.29 Community, on a single server (bookmarks are not passed along).

Not ported yet

LLM entity extraction, entity deduplication and LLM relation extraction (regex extraction only)
GraphRetriever (hybrid vector and graph retrieval)
Backfills and the migration runner
The Postgres audit writer (a no-op audit sink is the default)