Demonstrates ArcadeDB's multi-model capabilities by implementing a Graph RAG (Retrieval-Augmented Generation) system that unifies three retrieval signals in a single database:

• Graph traversal — multi-hop entity bridging via knowledge graph relationships
• Vector similarity — semantic chunk retrieval using embeddings
• Full-text indexing — keyword-based content lookup

• Docker and Docker Compose
• curl and jq
• Java 21+ and Maven 3.x (for the Java demos)

docker compose up -d

./setup.sh

This creates the GraphRAG database, applies the schema, and inserts sample data.

./queries/queries.sh

cd java
mvn package -q
java -jar target/graph-rag.jar

cd langchain4j
mvn package -q

# Embedding store: ingest + similarity search
java -jar target/graph-rag-langchain4j.jar

# Content retriever: semantic search + graph expansion
java -cp target/graph-rag-langchain4j.jar com.arcadedb.examples.GraphRAGContentRetriever

• 8 chunks from 4 internal documents with 4D embeddings
• 11 entities (4 persons, 4 concepts, 3 organizations)
• ~25 edges (MENTIONS, RELATES_TO, WORKS_AT, AUTHORED)
• Multi-hop design: querying "Vector Search" bridges to GraphRAG docs via shared entity mentions

The langchain4j/ directory contains two standalone examples using LangChain4j with ArcadeDB via the Neo4j Bolt protocol:

• GraphRAGEmbeddingStore — ingests text chunks with real 384D embeddings (AllMiniLmL6V2), stores via Cypher over Bolt, and performs similarity search using LangChain4j's cosine similarity
• GraphRAGContentRetriever — re-embeds the sample Chunk data with 384D vectors, runs semantic search, then enriches results with graph context via Cypher traversal (entities mentioned by top matches)

No external API keys required — the embedding model runs in-process.

This use case targets ArcadeDB 26.3.1. Vector similarity queries use vectorNeighbors('IndexName[property]', vector, k) with an LSM_VECTOR index. The Bolt protocol (port 7687) enables Neo4j driver compatibility.

ArcadeDB Graph RAG use case