Hey HN, we want to share HelixDB, a project we have been building to solve the 'multi-hop' problem in enterprise AI. Why a hybrid? Vector databases are powerful for similarity searches, but they struggle when you need to navigate the complex relationships—like who worked on a specific ticket, which meeting defined the scope, and which code commit addressed the issue. When your RAG application needs to connect these dots across thousands of documents, standard vector search often hits a wall. HelixDB is a native Graph-Vector database that solves this by keeping your semantic embeddings and structural relationships in one high-performance, ACID-compliant engine.

Use Cases for HelixDB

• Engineering Traceability: you can map a production bug back through the Jira ticket to the original design spec and the developer who submitted the PR, ensuring full visibility into the development lifecycle.
• Knowledge Management: you can construct an organizational 'memory' that links Slack conversations, technical documentation, and meeting transcripts, helping new hires find answers through semantic and relationship-based traversal.
• Risk & Compliance: you can identify clusters of risky relationships or unauthorized data access patterns by traversing deep-linked organizational hierarchies to ensure data governance.

Why a Native Graph-Vector Architecture?

Many people have a thing against “yet another database engine,” but we went ahead and did it anyway, because we think it makes working with our database so much easier. Stitching together a separate vector store with a graph database creates massive data movement latency and consistency nightmares. By using Rust for our core engine, we ensure that your graph traversals and vector lookups happen in the same memory space, reducing overhead significantly.

Performance Benchmarking

We are obsessive about speed. Our benchmarking shows that we are on par with Pinecone and Qdrant for vectors, and our graph is up to three orders of magnitude faster than Neo4j because we eliminate the need for cross-service network calls. Check out our detailed benchmarking report here: benchmarks.

Get Started

We are building this to be the backbone of agentic AI. If you’d like to try it out in a simple RAG demo, you can follow this guide: quickstart. We are actively looking for contributors and testers to stress-test our query DSL. Many thanks! Comments and feedback welcome!