HelixDB Tutorial: Building a Hybrid RAG Pipeline with Graph + Vector Search

In this hands-on walkthrough we build a hybrid RAG pipeline with HelixDB — combining graph traversal with vector similarity so a query like "find similar papers to X that also cite Y" returns genuinely useful results. This tutorial assumes basic familiarity with databases and Python; the full querying guide is in the official docs.

Step 1: Start HelixDB

HelixDB runs on object storage, so there is no local disk to provision. The quickest path is the container image built from the repository's Dockerfile:

# clone and build
git clone https://github.com/HelixDB/helix-db
cd helix-db && docker build -t helixdb .
# run pointing at your S3-compatible bucket
docker run -d -p 4242:4242 \
  -e HELIX_BUCKET=my-ai-data -e HELIX_ENDPOINT=https://s3.example.com \
  --name helixdb helixdb

Alternative: download the prebuilt Rust binary and run it with the same environment variables.

Step 2: Define Schema — Nodes, Edges, and Embeddings

HelixDB's single schema covers all three data models. Create a Paper node type with a title property, a cites edge type, and a embedding vector field:

CREATE NODE TYPE Paper {
  title: STRING,
  abstract: STRING,
  embedding: VECTOR(384)
};
CREATE EDGE TYPE cites FROM Paper TO Paper;

Step 3: Hybrid Query — Vector + Graph in One Call

The real power of HelixDB is querying both dimensions simultaneously:

SELECT p.title, p.abstract
FROM Paper p
MATCH (c)-[:cites]->(p)
WHERE c.title = 'Attention Is All You Need'
ORDER BY SIMILARITY(p.embedding, $query_vector) DESC
LIMIT 10;

This returns papers cited by the transformer paper, ranked by semantic similarity — a single round-trip instead of a vector lookup plus a graph join across two databases.

Step 4: What We Learned

HelixDB turns "graph database plus vector database" into one clean deployment. For teams building knowledge-graph assistants, agent memory, or recommendation systems, it is well worth a proof of concept. Ready to try it?